Multi-source information association system and method based on entity link in agricultural scene
By building a cross-modal knowledge graph and a low-code workflow engine, the integration and association problems of multimodal data in the agricultural management system are solved, efficient and automated agricultural knowledge processing is achieved, and the intelligent development of agriculture is promoted.
Patent Information
- Application Number
- CN202510700621.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-28
- Publication Date
- 2025-08-26
AI Technical Summary
The existing agricultural management systems have shortcomings in multimodal data compatibility, dynamic correlation capabilities and executability, making it difficult to efficiently integrate multi-source heterogeneous data and make intelligent decisions.
A multi-source information association system based on entity links is adopted to build a cross-modal knowledge graph through multimodal data analysis and entity linking technology, and a GraphRAG hierarchical search technology is used for accurate search, and an executable work plan is generated through a low-code workflow engine.
It realizes effective integration and correlation of different data sources, provides efficient and automated agricultural knowledge processing solutions, and improves the processing efficiency and intelligence level of agricultural knowledge.
Smart Images

Figure DJMQHRAU7OTLQ6DVSC1RQRDSDEM8OQZD7LUQ8S53 
Figure O6SRZ6DI8OCKDKI03OXYFCNPP288T3BWKHLJUHYV
Abstract
Description
Technical Field
[0001] The present invention relates to the field of agricultural artificial intelligence technology, and in particular to a multi-source information association system based on entity linking in agricultural scenarios. Background Art
[0002] In the field of smart agriculture, agricultural production management relies on multi-source, heterogeneous data (such as soil monitoring, weather forecasts, crop growth data, and agricultural machinery operation logs). Efficiently integrating multimodal data and enabling intelligent decision-making is a key challenge within the industry. Traditional agricultural management relies on expert experience or single-point data analysis, making it difficult to dynamically adapt to complex agricultural production environments.
[0003] Currently, some AI systems use knowledge graph technology to associate agricultural data. They propose using agricultural knowledge graphs to store crop growth rules and generate farming recommendations through rule engines. However, these solutions have the following core flaws: poor multimodal data compatibility, insufficient dynamic association capabilities, and limited feasibility.
[0004] Therefore, the present invention provides a multi-source information association system and method based on entity linking in an agricultural scenario. Summary of the Invention
[0005] This invention provides a multi-source information association system and method based on entity linking in agricultural scenarios. Through multimodal data parsing and entity linking technology, this system constructs a cross-modal knowledge graph, effectively integrating and linking different data sources in the agricultural field. This system enables accurate retrieval and optimization of agricultural knowledge, providing an efficient and automated solution for agricultural production. This system improves the efficiency of agricultural knowledge processing and promotes the development of intelligent agriculture.
[0006] The present invention provides a multi-source information association system based on entity linking in an agricultural scenario, comprising: Data parsing module: Builds a multimodal data parsing layer to perform data parsing. The parsed data is converted into knowledge units in a preset format through the Converter component, preserving the original semantic hierarchical structure and generating knowledge association anchors. Graph construction module: uses knowledge association anchors and entity linking technology to build a cross-modal knowledge graph; Graph retrieval module: retrieves cross-modal knowledge graphs based on GraphRAG hierarchical retrieval technology; Result optimization module: optimizes the search results to obtain optimized search results; Result processing module: Build a low-code workflow engine based on the optimized search results, and then build an agricultural knowledge processing pipeline, and execute the agricultural knowledge processing pipeline to obtain an executable work plan.
[0007] A multi-source information association system based on entity linking in an agricultural scenario, further comprising: Visualization module: renders optimized search results into interactive knowledge cards; Feedback module: collects users' accuracy scores on executable work plans; Incremental learning module: Optimizes the cross-modal knowledge graph based on users' accuracy scores of executable work solutions.
[0008] Preferably, the data parsing module comprises: a text parsing unit: extracting entities from agricultural literature based on a text parser, outputting structured text data and performing semantic analysis; Visual parsing unit: This unit uses a visual parser to identify agricultural machinery and operating actions in agricultural operation videos and generates spatiotemporal labeled data. Time series parsing unit: processes meteorological sensor time series data based on the time series parser and detects abnormal fluctuation fragment data; Format conversion unit: Converts the semantic analysis results, spatiotemporal tag data, and abnormal fluctuation fragment data into knowledge units with preset formats and hyperlinks through the Converter component; Anchor point generation unit: extracts TF-IDF keywords from the converted knowledge units and generates cross-modal associated anchor points.
[0009] Preferably, the graph construction module includes: an entity linking unit: constructing a four-layer disambiguation system to eliminate synonymous entity ambiguities in knowledge association anchors, and then aligning all knowledge association anchors to obtain a number of entity nodes; Relationship extraction unit: LLM with few-shot hints is used to extract semantic relationships between entity nodes from knowledge association anchors; Community division unit: Apply the Leiden algorithm to cluster entity nodes to form a mutually exclusive community structure, where the mutually exclusive community structure contains several communities; Summary generation unit: Generates a community summary represented by an embedding for each community; Graph structure generation unit: Constructs a cross-modal knowledge graph with a weighted directed graph structure based on entity nodes, semantic relationships between entity nodes, and community summaries of each community.
[0010] Preferably, the entity link unit includes: System construction subunit: Based on the preset agricultural domain ontology library, a four-layer disambiguation system is constructed, including: word surface layer, context layer, modality layer, and spatiotemporal layer; Disambiguation subunit: The word-face layer uses the improved Jaro-Winkler algorithm to match entity aliases. The context layer uses a fine-tuned RoBERTa model to analyze surrounding entities and determine referential relationships. The modality layer calculates the CLIP embedding similarity of cross-media entities. The spatiotemporal layer uses GPS trajectories and time windows to filter conflicting entities. Anchor point alignment subunit: Generate a globally unique URI identifier based on the disambiguated entity, establish an association with the preset standard vocabulary, and then obtain the aligned associated anchor points to obtain several entity nodes.
[0011] Preferably, the graph retrieval module includes: Query parsing unit: identifies core entities based on the cross-modal knowledge graph, expands the search scope along the semantic relationship path of the core entities, and generates candidate search subgraphs; Hierarchical retrieval unit: uses a global-local hybrid retrieval strategy to perform retrieval in candidate subgraphs; Result integration unit: integrates the retrieval results of the global and local stages and fuses them into structured retrieval results.
[0012] Preferably, the result optimization module includes: Confidence evaluation unit: Confidence scoring of search results, verification of semantic consistency between knowledge unit and knowledge association anchor based on preset knowledge unit format, and screening of high-confidence candidate result sets; Multi-source evidence fusion unit: associates and compares high-confidence candidate results with knowledge units, aligns cross-modal data through knowledge association anchors, eliminates result conflicts, and completes missing semantic information to obtain an optimized high-confidence candidate result set; Result optimization unit: Based on the structured relationships of the cross-modal knowledge graph, the logical consistency scores of entities and relationships in the high-confidence candidate result set are determined, and contradictory nodes that violate domain rules are eliminated to obtain the final high-confidence candidate result set; Weight adjustment unit: Dynamically adjusts the association weight of the final high-confidence candidate result set in the cross-modal knowledge graph based on the logical consistency score, and outputs a result list with weight annotations; Result adjustment unit: Feeds back the result list with weight annotation to the graph retrieval module, triggers a secondary search based on the new weights, forms a closed-loop optimization mechanism, and finally outputs the optimized retrieval results.
[0013] Preferably, the result processing module includes: Process configuration unit: configures executable task processes based on optimized search results; Business adaptation unit: maps optimized search results to predefined business logic templates based on preset agricultural decision rules; Permission verification unit: obtains user identity authentication information and controls the execution authority of the task process based on the user identity authentication information; Service execution unit: calls related calculation and retrieval services in sequence according to the task flow to generate the final decision plan; Pipeline construction unit: Integrate the above-mentioned optimized retrieval, business adaptation and service execution processes to form an agricultural knowledge processing pipeline.
[0014] Solution acquisition unit: executes the agricultural knowledge processing pipeline to obtain an executable work solution.
[0015] A multi-source information association method based on entity linking in agricultural scenarios, comprising: Step 1: Build a multimodal data parsing layer to perform data parsing. The parsed data is converted into knowledge units in a preset format through the Converter component, preserving the original semantic hierarchical structure and generating knowledge association anchors. Step 2: Use knowledge association anchors and entity linking technology to build a cross-modal knowledge graph; Step 3: Search the cross-modal knowledge graph based on GraphRAG hierarchical retrieval technology; Step 4: Optimize the search results to obtain optimized search results; Step 5: Build a low-code workflow engine based on the optimized search results, and then build an agricultural knowledge processing pipeline. Execute the agricultural knowledge processing pipeline to obtain an executable work plan.
[0016] Compared with the prior art, the present invention has the following advantages: By using multimodal data parsing and entity linking technology to construct a cross-modal knowledge graph, the system effectively integrates and links different agricultural data sources, accurately retrieves and optimizes agricultural knowledge, and provides efficient and automated solutions for agricultural production. This system improves the efficiency of agricultural knowledge processing and promotes the development of intelligent agriculture. BRIEF DESCRIPTION OF THE DRAWINGS
[0017] In order to more clearly illustrate the technical solutions in the present invention or the prior art, a brief introduction is given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.
[0018] Figure 1 This is a structural diagram of a multi-source information association system based on entity linking in an agricultural scenario provided by an embodiment of the present invention.
[0019] Figure 2This is a flow chart of a multi-source information association method based on entity linking in an agricultural scenario provided by an embodiment of the present invention. DETAILED DESCRIPTION
[0020] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the embodiments described are only some of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.
[0021] Example 1: The embodiment of the present invention provides a multi-source information association system based on entity linking in an agricultural scenario. Figure 1 Shown, including: Data parsing module: Builds a multimodal data parsing layer to perform data parsing. The parsed data is converted into knowledge units in a preset format through the Converter component, preserving the original semantic hierarchical structure and generating knowledge association anchors. Graph construction module: uses knowledge association anchors and entity linking technology to build a cross-modal knowledge graph; Graph retrieval module: retrieves cross-modal knowledge graphs based on GraphRAG hierarchical retrieval technology; Result optimization module: optimizes the search results to obtain optimized search results; Result processing module: Build a low-code workflow engine based on the optimized search results, and then build an agricultural knowledge processing pipeline, and execute the agricultural knowledge processing pipeline to obtain an executable work plan.
[0022] In this example, GraphRAG (Graph-based Retrieval-Augmented Generation) is an AI technology that combines knowledge graphs with hierarchical retrieval, enabling efficient queries through a multi-level index structure. The "hierarchical" nature of GraphRAG refers to a two-step retrieval process: global to local. Coarse-grained retrieval rapidly locates candidate subgraphs based on the knowledge graph's topological structure (e.g., entity categories and relationship density). Fine-grained refined ranking uses semantic embeddings (e.g., BERT vectors) within the subgraph to match the user's query intent and generate precise answers. For an agricultural example, when a user queries "corn rust prevention and control methods," the coarse search first targets the "pest and disease control" subgraph (including nodes such as diseases, pesticides, and crops). Refined ranking then selects nodes related to "corn rust" (e.g., "triadimefon spray" and "temperature threshold for disease onset"), ultimately outputting a prevention and control plan.
[0023] In this embodiment, the low-code workflow engine is a tool for configuring business processes through a graphical interface (non-programming). Users can build automated pipelines by dragging and dropping predefined components (such as conditional branches and API calls). Its core features include: visual orchestration: converting logical relationships into flow charts (such as "If soil moisture < 30% → Start irrigation"); modular reuse: built-in agricultural templates (such as the "Fertilization Work Order Generator") can be directly called. Agricultural example: Build an "intelligent irrigation" workflow: drag the "Soil Sensor" node to bind real-time data; add the "Conditional Judgment" node to set the threshold (humidity < 30%); connect the "Agricultural Machinery Control" node to trigger irrigation instructions, without writing any code.
[0024] The beneficial effects of this technical solution include: building a cross-modal knowledge graph through multimodal data parsing and entity linking technology, effectively integrating and linking diverse agricultural data sources, accurately retrieving and optimizing agricultural knowledge, and providing efficient and automated solutions for agricultural production. This system improves the efficiency of agricultural knowledge processing and promotes the development of intelligent agriculture.
[0025] Example 2: An embodiment of the present invention provides a multi-source information association system based on entity linking in an agricultural scenario, further comprising: Visualization module: renders optimized search results into interactive knowledge cards; Feedback module: collects users' accuracy scores on executable work plans; Incremental learning module: Optimizes the cross-modal knowledge graph based on users' accuracy scores of executable work solutions.
[0026] In this example, interactive knowledge cards are dynamic visualization components that encapsulate graph search results (such as entity attributes and relationships) into clickable, filterable graphical units, allowing users to explore multidimensional information through interactions (such as hover highlights and drop-down menus). An agricultural example displays a "wheat fusarium head blight" card, including an image of the pathogen, a list of control agents (click to view suppliers), and a climate risk map (with a sliding threshold). Users can directly select the agent to generate a purchase order.
[0027] In this embodiment, the accuracy score is a user's quantitative evaluation of the actual effect of the system-generated plan (such as 1-5 stars), which is collected through the feedback interface and is used to measure the degree of match between AI decisions and real-world scenario requirements. For example, in agriculture, after farmers implement the "soybean fertilization plan", they score 4 stars based on the yield improvement rate. The system records this score and stores it in association with the plan parameters (such as nitrogen-phosphorus ratio and fertilization timing).
[0028] In this embodiment, the optimization graph uses scoring data to adjust the entity weights or association confidences in the graph. Low-scoring solutions trigger reverse reasoning to correct incorrect nodes or supplement missing relationships. For example, in agriculture, multiple users gave low scores to the "tomato late blight" treatment plan. The system traced the source and found that the "pathogen resistance" attribute in the graph was expired. It was automatically updated to the 2024 regional resistance monitoring data and strengthened the association with new agents.
[0029] The beneficial effects of the above technical solution include further optimizing the entity-linking-based multi-source information association AI system for agricultural scenarios through visualization, feedback, and incremental learning modules. The visualization module presents optimized search results as interactive knowledge cards, making them easier for users to understand and operate. The feedback module collects user ratings of work plans, and the incremental learning module continuously optimizes the cross-modal knowledge graph based on user feedback, enhancing the system's intelligence and personalized response capabilities, thereby improving the efficiency and accuracy of agricultural knowledge processing.
[0030] Example 3: The embodiment of the present invention provides a multi-source information association system based on entity linking in an agricultural scenario, and a data parsing module, comprising: a text parsing unit: extracting entities from agricultural literature based on a text parser, outputting structured text data and performing semantic analysis; Visual parsing unit: This unit uses a visual parser to identify agricultural machinery and operating actions in agricultural operation videos and generates spatiotemporal labeled data. Time series parsing unit: processes meteorological sensor time series data based on the time series parser and detects abnormal fluctuation fragment data; Format conversion unit: Converts the semantic analysis results, spatiotemporal tag data, and abnormal fluctuation fragment data into knowledge units with preset formats and hyperlinks through the Converter component; Anchor point generation unit: extracts TF-IDF keywords from the converted knowledge units and generates cross-modal associated anchor points.
[0031] In this embodiment, the text parsing unit includes: a PDF structure analysis subunit: parsing the document directory hierarchy to generate a tree index; a table extraction subunit: using TableTransformer to identify and escape table data; a term standardization subunit: matching the unified expression form of the agricultural terminology list; a context retention subunit: maintaining the reference association in the original text; and a quality verification subunit: checking data integrity through a rule engine.
[0032] In this example, the visual parser (CV model) analyzes agricultural videos, identifies agricultural machinery (such as a seed drill) and actions (such as fertilizing), and generates labeled data with timestamps and coordinates. For example, the third frame of the labeled video shows "tractor subsoiling" and the location is farmland area A.
[0033] In this embodiment, the time series analyzer uses a signal processing algorithm to analyze meteorological data (temperature, humidity), detect anomalies (e.g., humidity > 90% for three consecutive hours), and output an abnormal period marker. For example, it identifies a sudden increase in soil moisture caused by a sudden downpour between 2:00 PM and 4:00 PM on May 10th.
[0034] In this example, the Converter component, a format standardization module, converts multi-source data (text entities, video tags, and weather clips) into unified knowledge units (JSON-LD format), embedding hyperlinks to associate graph nodes. For example, a "rust prevention" text entity is hyperlinked to a related video clip.
[0035] In this example, the association anchor is a cross-modal search keyword (a high-TF-IDF weighted keyword) that serves as a hub for linking text, video, and time series data. For example, the keyword "spring irrigation" is associated with a text plan, agricultural machinery videos, and soil moisture data.
[0036] The beneficial effects of this technical solution include: The data parsing module improves the data processing capabilities of the entity-linking-based multi-source information association AI system in agricultural scenarios. The text parsing unit extracts entities from agricultural literature and performs semantic analysis; the visual parsing unit identifies agricultural machinery and operating actions from agricultural operation videos; and the time series parsing unit processes meteorological sensor time series data and detects abnormal fluctuations. The format conversion unit uniformly converts data into knowledge units, and the anchor point generation unit extracts keywords and generates cross-modal association anchors, ensuring the effective integration of multi-source information and supporting precise agricultural knowledge association and application.
[0037] Example 4: The embodiment of the present invention provides a multi-source information association system based on entity linking in an agricultural scenario, and a graph construction module, including: an entity linking unit: constructing a four-layer disambiguation system to eliminate synonymous entity ambiguities in knowledge association anchor points, and then aligning all knowledge association anchor points to obtain multiple entity nodes; Relationship extraction unit: LLM with few-shot hints is used to extract semantic relationships between entity nodes from knowledge association anchors; Community division unit: Apply the Leiden algorithm to cluster entity nodes to form a mutually exclusive community structure, where the mutually exclusive community structure contains several communities; Summary generation unit: Generates a community summary represented by an embedding for each community; Graph structure generation unit: Constructs a cross-modal knowledge graph with a weighted directed graph structure based on entity nodes, semantic relationships between entity nodes, and community summaries of each community.
[0038] In this example, a few-shot LLM with prompts is used to extract semantic relationships between entity nodes from knowledge association anchors. This involves designing a hybrid relationship extraction framework. For explicit relationships (e.g., the direct statement "nitrogen fertilizer promotes tillering" in a document), a prompt-tuning LLM (e.g., ChatGLM3) is used to generate <subject, predicate, object> triples. For implicit relationships (e.g., the causal chain "continuous rainfall → increased humidity → disease risk" in sensor data), a temporal logic network is used for probabilistic reasoning. Specifically, a visual relationship detection module is developed for image-text cross-modal associations, mapping the actions <agricultural machinery, spraying, farmland> detected by YOLOv8 to SWRL rule expressions. All relationships are validated using a Bayesian network, and only those with ≥3 supporting evidence are stored to avoid misassociations caused by data sparsity.
[0039] In this embodiment, Few-Shot Prompting in LLMs is simply explained as follows: a large language model (LLM) learns task rules through a small number of examples (usually 1-10), rather than training from scratch or performing zero-example reasoning. The core principle is to provide a task description + a small number of input-output examples in the input (such as "Translation: apple → apple; banana → banana"), and the LLM quickly summarizes the pattern through the examples and then processes new input.
[0040] In this embodiment, the Leiden algorithm: a modularity-optimized community discovery algorithm, clusters entity nodes into non-overlapping communities (such as "pest control community" and "agricultural machinery operation community") according to relationship density, thereby improving the interpretability of the graph.
[0041] In this embodiment, a community summary, represented by an embedded representation, is generated for each community. Two types of summaries are generated for each community: a structured summary using the OWL rdfs:comment property to describe the community topic; and a vectorized summary generated through GraphSAGE to generate a 128-dimensional community embedding for subsequent coarse screening in the retrieval module. A dynamic adjustment mechanism is implemented for community boundaries, triggering reconstruction when the association strength of newly added entities exceeds a threshold. Community summaries use graph embeddings (e.g., Node2Vec) to compress entity relationships within the community into low-dimensional vector representations, capturing the core semantics (e.g., "Community 1 Summary: Soil Improvement Technology and Related Agricultural Machinery").
[0042] In this embodiment, the weighted directed graph structure is a storage form of the knowledge graph, where nodes are entities, edges are semantic relationships (direction represents subject and object), and weights represent the strength of the relationship (such as "fertilizing → promoting growth" has a weight of 0.9).
[0043] The beneficial effects of the above technical solution are: The graph construction module enhances the knowledge graph construction capabilities of the multi-source information association AI system based on entity linking in agricultural scenarios. The entity linking unit eliminates the ambiguity of synonymous entities and aligns knowledge association anchors. The relationship extraction unit uses the LLM suggested by a few samples to extract semantic relationships between entities. The community division unit applies the Leiden algorithm to cluster entity nodes and generate mutually exclusive community structures. The summary generation unit generates a summary for each community, and the graph structure generation unit constructs a cross-modal knowledge graph with a weighted directed graph. This solution effectively improves the accuracy and usability of the knowledge graph and supports precise agricultural data analysis and application.
[0044] Example 5: An embodiment of the present invention provides a multi-source information association system based on entity linking in an agricultural scenario, wherein the entity linking unit includes: System construction subunit: Based on the preset agricultural domain ontology library, a four-layer disambiguation system is constructed, including: word surface layer, context layer, modality layer, and spatiotemporal layer; Disambiguation subunit: The word-face layer uses the improved Jaro-Winkler algorithm to match entity aliases. The context layer uses a fine-tuned RoBERTa model to analyze surrounding entities and determine referential relationships. The modality layer calculates the CLIP embedding similarity of cross-media entities. The spatiotemporal layer uses GPS trajectories and time windows to filter conflicting entities. Anchor point alignment subunit: Generate a globally unique URI identifier based on the disambiguated entity, establish an association with the preset standard vocabulary, and then obtain the aligned associated anchor points to obtain several entity nodes.
[0045] In this embodiment, the context layer uses a fine-tuned RoBERTa model to analyze surrounding entities to determine referential relationships. For example, in the "early rice" paragraph, "it" refers to rice seeds rather than land. In this embodiment, the conflicting entities are because it is impossible for harvesting and sowing operations to occur simultaneously on the same plot of land.
[0046] In this embodiment, the word layer: uses the improved Jaro-Winkler algorithm (increasing the weight of agricultural affixes) to match aliases (such as "potato" vs. "potato"); the context layer: based on a fine-tuned version of the RoBERTa model, the referent target is inferred through contextual entities (such as "planting" and "disease"); the modal layer: across text / image / video data, the CLIP model is used to calculate the embedding similarity of different modal entities (such as matching the text description of "tractor" with the picture of agricultural machinery); the spatiotemporal layer: combines GPS tracks (such as farmland coordinates) and time windows (crop growth cycles) to eliminate logically conflicting entities (such as "harvesting" and "sowing" do not occur at the same time).
[0047] In this embodiment, a global URI identifier is used to assign a unique resource identifier (such as agri:maize-crop#2023) to the disambiguated entity, thereby achieving standardized mapping of multi-source data.
[0048] The beneficial effects of the above technical solution are: the entity linking unit is used to improve the entity disambiguation and alignment capabilities of the multi-source information association AI system based on entity linking in agricultural scenarios. The system construction subunit constructs a four-layer meta-system based on the agricultural field ontology library, including the word layer, context layer, modal layer and spatiotemporal layer. The ambiguity elimination subunit combines the improved Jaro-Winkler algorithm, the fine-tuned RoBERTa model, the CLIP embedding similarity and the GPS trajectory to achieve multi-dimensional entity disambiguation. The anchor point alignment subunit generates a globally unique URI identifier and aligns it with the standard vocabulary to ensure the accuracy and consistency of the entity node, providing high-quality basic data for the cross-modal knowledge graph.
[0049] Example 6: The embodiment of the present invention provides a multi-source information association system based on entity linking in an agricultural scenario, and a graph retrieval module, including: Query parsing unit: identifies core entities based on the cross-modal knowledge graph, expands the search scope along the semantic relationship path of the core entities, and generates candidate search subgraphs; Hierarchical retrieval unit: uses a global-local hybrid retrieval strategy to perform retrieval in candidate subgraphs; Result integration unit: integrates the retrieval results of the global and local stages and fuses them into structured retrieval results.
[0050] In this embodiment, a global-local hybrid retrieval strategy is used to perform retrieval in the candidate subgraph: global retrieval stage: random block retrieval is performed based on community summaries to generate preliminary candidate answers; local retrieval stage: query semantics are optimized through embedding vector matching to screen out exactly matching entities and relationships.
[0051] In this embodiment, core entity recognition and semantic path expansion: when parsing user queries (such as "organic methods for preventing and controlling wheat fusarium wilt"), the key entity ("wheat fusarium wilt") is first located, and then the related entities are expanded along the graph relationship ("pathogens → prevention and control measures → organic pesticides") to generate candidate subgraphs (including nodes such as "antagonistic bacteria" and "biological pesticides").
[0052] The beneficial effects of the above technical solution are: the graph retrieval module improves the retrieval efficiency and accuracy of the multi-source information association AI system based on entity linking in agricultural scenarios. The query parsing unit identifies core entities based on the cross-modal knowledge graph and expands the search scope along the semantic relationship path to generate candidate subgraphs. The hierarchical retrieval unit uses a global-local hybrid retrieval strategy to perform retrieval and optimize the retrieval results. The result integration unit merges the retrieval results of the global and local stages into structured results, improving the comprehensiveness and accuracy of the retrieval and ensuring the rapid and accurate acquisition of relevant agricultural knowledge information.
[0053] Example 7: The embodiment of the present invention provides a multi-source information association system based on entity linking in an agricultural scenario, and a result optimization module, including: Confidence evaluation unit: Confidence scoring of search results, verification of semantic consistency between knowledge unit and knowledge association anchor based on preset knowledge unit format, and screening of high-confidence candidate result sets; Multi-source evidence fusion unit: associates and compares high-confidence candidate results with knowledge units, aligns cross-modal data through knowledge association anchors, eliminates result conflicts, and completes missing semantic information to obtain an optimized high-confidence candidate result set; Result optimization unit: Based on the structured relationships of the cross-modal knowledge graph, the logical consistency scores of entities and relationships in the high-confidence candidate result set are determined, and contradictory nodes that violate domain rules are eliminated to obtain the final high-confidence candidate result set; Weight adjustment unit: Dynamically adjusts the association weight of the final high-confidence candidate result set in the cross-modal knowledge graph based on the logical consistency score, and outputs a result list with weight annotations; Result adjustment unit: Feeds back the result list with weight annotation to the graph retrieval module, triggers a secondary search based on the new weights, forms a closed-loop optimization mechanism, and finally outputs the optimized retrieval results.
[0054] In this embodiment, the confidence assessment unit includes: knowledge unit format verification: checking whether the search results conform to the preset agricultural knowledge structure (such as the "disease-control method-pesticide" triple) to ensure semantic correctness, such as excluding invalid associations such as "wheat fusarium -> corn planting"; semantic consistency scoring: calculating the matching degree between candidate results and knowledge anchors (standard entity URIs) based on a pre-trained model (such as BERT), and filtering low-scoring items (such as similarity <0.7).
[0055] In this embodiment, the multi-source evidence fusion unit: cross-modal alignment: using knowledge association anchors (global URIs) to associate data from different sources (such as a text report "using Bacillus subtilis" + an experimental video "the effect of spraying microbial agents") to fill information gaps (such as supplementing the concentration of pesticide application); conflict resolution: eliminating contradictory evidence through spatiotemporal layers (such as sensor data from the same farmland) or domain rules (such as "organic pesticides cannot contain chemical synthetic ingredients").
[0056] In this embodiment, the result optimization unit: logical consistency score: based on the topological structure of the knowledge graph (such as the "pathogen-host-control" chain), verifies the reasonable connection of entities (such as the host of "wheat fusarium" must be a gramineous crop), and deletes illegal nodes (such as those associated with "fruit trees").
[0057] In this embodiment, the weight adjustment unit increases the edge weight of the reliable result according to the consistency score (for example, the weight of "biological control → effectiveness" is increased from 0.6 to 0.9), which affects the priority of subsequent retrieval.
[0058] In this embodiment, the closed-loop optimization mechanism includes: feedback secondary search: feeding back the weighted updated results to the graph search module, re-triggering the query (such as prioritizing the search for high-weight "organic prevention" paths), and iteratively optimizing the results.
[0059] The beneficial effects of this technical solution include: The result optimization module improves the retrieval quality of the entity-linking-based multi-source information association AI system in agricultural scenarios. The confidence assessment unit scores retrieval results and selects high-confidence candidate results. The multi-source evidence fusion unit aligns cross-modal data to eliminate conflicts and complete missing semantic information. The result optimization unit eliminates conflicting nodes based on logical consistency scores to ensure that the results comply with domain rules. The weight adjustment unit dynamically adjusts the weight of the final result in the knowledge graph, forming a closed-loop optimization mechanism that ultimately outputs precise optimized retrieval results, improving the reliability and accuracy of the system.
[0060] Example 8: The embodiment of the present invention provides a multi-source information association system based on entity linking in an agricultural scenario, and a result processing module, including: Process configuration unit: configures executable task processes based on optimized search results; Business adaptation unit: maps optimized search results to predefined business logic templates based on preset agricultural decision rules; Permission verification unit: obtains user identity authentication information and controls the execution authority of the task process based on the user identity authentication information; Service execution unit: calls related calculation and retrieval services in sequence according to the task flow to generate the final decision plan; Pipeline construction unit: Integrate the above-mentioned optimized retrieval, business adaptation and service execution processes to form an agricultural knowledge processing pipeline.
[0061] Solution acquisition unit: executes the agricultural knowledge processing pipeline to obtain an executable work solution.
[0062] In this embodiment, the process configuration unit: low-code interface layer: provides a visual drag-and-drop component library (such as task nodes, conditional branches, loop controllers), and maps the logical relationships in the optimized search results (such as "if soil moisture <30%, trigger irrigation tasks") into editable workflows; dynamic parameter injection: allows users to directly bind variables in the search results through a form (such as "irrigation amount = soil water demand calculation results") to generate parameterized task processes.
[0063] In this embodiment, the business adaptation unit: Template matching engine: automatically matches predefined business logic templates (such as "water and fertilizer management template for wheat greening period") based on the agricultural decision rule library (such as crop growth model and agricultural machinery scheduling strategy), and fills the search results into the template variable slots; Cross-modal adapter: supports converting unstructured search results (such as pest and disease image classification labels) into structured business instructions (such as "spray 5% imidacloprid solution"); In this embodiment, the permission verification unit: Role-resource bidirectional verification: through the user identity authentication information associated with the agricultural machinery operation permission database (for example, "Farmer A can only operate the sprinkler equipment bound to his own field"), the illegal operation nodes in the task process are dynamically trimmed; In this embodiment, the service execution unit: a microservice orchestrator: calls distributed services (such as weather forecast services and agricultural machinery path planning services) according to task processes, using asynchronous message queues (such as RabbitMQ) to ensure execution reliability under high concurrency; real-time feedback integration: embeds a sensor data monitoring submodule during the execution process to dynamically modify decision plans (such as "rainfall detected during irrigation task execution, automatically terminate the task"); In this embodiment, the pipeline construction unit: a logical breakpoint resumption mechanism: if the pipeline execution is interrupted (such as a network failure), a snapshot of the current node status is recorded, and the completed steps are automatically skipped during recovery; a resource occupancy predictor: predicts the pipeline computing load based on historical data, and dynamically allocates cloud computing resources (such as AWS Lambda function instances), and also includes: a task scheduling subunit: dynamically allocates computing resources according to the priority of agricultural operations to ensure that high-timeliness tasks are executed first; a service dependency subunit: automatically parses the input and output dependencies between standardized interfaces and generates a directed acyclic execution graph; a consistency check subunit: verifies the consistency of data types and business rules when combining interfaces, and enforces constraints on agricultural field logic; a parallel optimization subunit: starts multi-threaded concurrent processing for retrieval and computing services without dependencies.
[0064] In this embodiment, the solution acquisition unit: multimodal output generator: converts the pipeline execution results into: machine executable instructions (such as agricultural machinery control instructions of ISO11783 standard); human-readable work orders (including graphic operation guides and QR code jump links), and also includes: decision deconstruction sub-unit: decomposes the multi-source correlation results output by the pipeline into atomic operation instructions for the agricultural production link; time constraint sub-unit: based on the timeliness requirements of the crop growth stage, attaches a time window label to each operation instruction; resource binding sub-unit: matches the agricultural machinery equipment status library and dynamically allocates instructions to agricultural machinery resources; fault-tolerant revision sub-unit: when a conflict between instructions and real-time sensor data is detected, triggers local pipeline reconstruction; visual delivery sub-unit: generates structured work orders that support loading of agricultural APPs, including operation path navigation and early warning prompt layers.
[0065] The beneficial effects of the above technical solution are: the decision support capabilities of the multi-source information association AI system based on entity linking in agricultural scenarios are enhanced through the result processing module. The process configuration unit configures the task process based on the optimized retrieval results, and the business adaptation unit maps the retrieval results to the agricultural decision rule template. The permission verification unit controls the execution authority of the task process to ensure security. The service execution unit sequentially calls the calculation and retrieval services to generate a decision plan. The pipeline construction unit integrates the optimized retrieval, business adaptation and service execution processes to form an agricultural knowledge processing pipeline, and ultimately outputs an executable work plan to improve decision-making efficiency and accuracy.
[0066] Example 9: The embodiment of the present invention provides a multi-source information association method based on entity linking in an agricultural scenario. Figure 2 Shown, including: Step 1: Build a multimodal data parsing layer to perform data parsing. The parsed data is converted into knowledge units in a preset format through the Converter component, preserving the original semantic hierarchical structure and generating knowledge association anchors. Step 2: Use knowledge association anchors and entity linking technology to build a cross-modal knowledge graph; Step 3: Search the cross-modal knowledge graph based on GraphRAG hierarchical retrieval technology; Step 4: Optimize the search results to obtain optimized search results; Step 5: Build a low-code workflow engine based on the optimized search results, and then build an agricultural knowledge processing pipeline. Execute the agricultural knowledge processing pipeline to obtain an executable work plan.
[0067] The beneficial effects of this technical solution include: building a cross-modal knowledge graph through multimodal data parsing and entity linking technology, effectively integrating and linking diverse agricultural data sources, accurately retrieving and optimizing agricultural knowledge, and providing efficient and automated solutions for agricultural production. This system improves the efficiency of agricultural knowledge processing and promotes the development of intelligent agriculture.
[0068] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.
Claims
1. A multi-source information association system based on entity linking in agricultural scenarios, characterized by: include: Data parsing module: Builds a multimodal data parsing layer to perform data parsing. The parsed data is converted into knowledge units in a preset format through the Converter component, preserving the original semantic hierarchical structure and generating knowledge association anchors. Graph construction module: uses knowledge association anchors and entity linking technology to build a cross-modal knowledge graph; Graph retrieval module: retrieves cross-modal knowledge graphs based on GraphRAG hierarchical retrieval technology; Result optimization module: optimizes the search results to obtain optimized search results; Result processing module: Build a low-code workflow engine based on the optimized search results, and then build an agricultural knowledge processing pipeline, and execute the agricultural knowledge processing pipeline to obtain an executable work plan.
2. The multi-source information association system based on entity linking in an agricultural scenario according to claim 1 is characterized in that: Also includes: Visualization module: renders optimized search results into interactive knowledge cards; Feedback module: collects users' accuracy scores on executable work plans; Incremental learning module: Optimizes the cross-modal knowledge graph based on users' accuracy scores of executable work solutions.
3. The multi-source information association system based on entity linking in an agricultural scenario according to claim 1 is characterized in that: The data parsing module includes: a text parsing unit that extracts entities from agricultural literature based on a text parser, outputs structured text data, and performs semantic analysis; Visual parsing unit: This unit uses a visual parser to identify agricultural machinery and operating actions in agricultural operation videos and generates spatiotemporal labeled data. Time series parsing unit: processes meteorological sensor time series data based on the time series parser and detects abnormal fluctuation fragment data; Format conversion unit: Converts the semantic analysis results, spatiotemporal tag data, and abnormal fluctuation fragment data into knowledge units with preset formats and hyperlinks through the Converter component; Anchor point generation unit: extracts TF-IDF keywords from the converted knowledge units and generates cross-modal associated anchor points.
4. The multi-source information association system based on entity linking in an agricultural scenario according to claim 3 is characterized in that: The graph construction module includes: Entity linking unit: Constructing a four-layer disambiguation system to eliminate synonymous entity ambiguities in knowledge association anchors, and then aligning all knowledge association anchors to obtain several entity nodes; Relationship extraction unit: LLM with few-shot hints is used to extract semantic relationships between entity nodes from knowledge association anchors; Community division unit: Apply the Leiden algorithm to cluster entity nodes to form a mutually exclusive community structure, where the mutually exclusive community structure contains several communities; Summary generation unit: Generates a community summary represented by an embedding for each community; Graph structure generation unit: Constructs a cross-modal knowledge graph with a weighted directed graph structure based on entity nodes, semantic relationships between entity nodes, and community summaries of each community.
5. The multi-source information association system based on entity linking in an agricultural scenario according to claim 4 is characterized in that: Entity link unit, including: System construction subunit: Based on the preset agricultural domain ontology library, a four-layer disambiguation system is constructed, including: word surface layer, context layer, modality layer, and spatiotemporal layer; Disambiguation subunit: The word-face layer uses the improved Jaro-Winkler algorithm to match entity aliases. The context layer uses a fine-tuned RoBERTa model to analyze surrounding entities and determine referential relationships. The modality layer calculates the CLIP embedding similarity of cross-media entities. The spatiotemporal layer uses GPS trajectories and time windows to filter conflicting entities. Anchor point alignment subunit: Generate a globally unique URI identifier based on the disambiguated entity, establish an association with the preset standard vocabulary, and then obtain the aligned associated anchor points to obtain several entity nodes.
6. The multi-source information association system based on entity linking in an agricultural scenario according to claim 1 is characterized in that: Graph retrieval module, including: Query parsing unit: identifies core entities based on the cross-modal knowledge graph, expands the search scope along the semantic relationship path of the core entities, and generates candidate search subgraphs; Hierarchical retrieval unit: uses a global-local hybrid retrieval strategy to perform retrieval in candidate subgraphs; Result integration unit: integrates the retrieval results of the global and local stages and fuses them into structured retrieval results.
7. The multi-source information association system based on entity linking in an agricultural scenario according to claim 1 is characterized in that: Result optimization module, including: Confidence evaluation unit: Confidence scoring of search results, verification of semantic consistency between knowledge unit and knowledge association anchor based on preset knowledge unit format, and screening of high-confidence candidate result sets; Multi-source evidence fusion unit: associates and compares high-confidence candidate results with knowledge units, aligns cross-modal data through knowledge association anchors, eliminates result conflicts, and completes missing semantic information to obtain an optimized high-confidence candidate result set; Result optimization unit: Based on the structured relationships of the cross-modal knowledge graph, the logical consistency scores of entities and relationships in the high-confidence candidate result set are determined, and contradictory nodes that violate domain rules are eliminated to obtain the final high-confidence candidate result set; Weight adjustment unit: Dynamically adjusts the association weight of the final high-confidence candidate result set in the cross-modal knowledge graph based on the logical consistency score, and outputs a result list with weight annotations; Result adjustment unit: Feeds back the result list with weight annotation to the graph retrieval module, triggers a secondary search based on the new weights, forms a closed-loop optimization mechanism, and finally outputs the optimized retrieval results.
8. The multi-source information association system based on entity linking in an agricultural scenario according to claim 1 is characterized in that: Result processing module, including: Process configuration unit: configures executable task processes based on optimized search results; Business adaptation unit: maps optimized search results to predefined business logic templates based on preset agricultural decision rules; Permission verification unit: obtains user identity authentication information and controls the execution authority of the task process based on the user identity authentication information; Service execution unit: calls related calculation and retrieval services in sequence according to the task flow to generate the final decision plan; Pipeline construction unit: Integrates the above-mentioned optimized retrieval, business adaptation and service execution processes to form an agricultural knowledge processing pipeline; Solution acquisition unit: executes the agricultural knowledge processing pipeline to obtain an executable work solution.
9. A multi-source information association method based on entity linking in agricultural scenarios, characterized by: include: Step 1: Build a multimodal data parsing layer to perform data parsing. The parsed data is converted into knowledge units in a preset format through the Converter component, preserving the original semantic hierarchical structure and generating knowledge association anchors. Step 2: Use knowledge association anchors and entity linking technology to build a cross-modal knowledge graph; Step 3: Search the cross-modal knowledge graph based on GraphRAG hierarchical retrieval technology; Step 4: Optimize the search results to obtain optimized search results; Step 5: Build a low-code workflow engine based on the optimized search results, and then build an agricultural knowledge processing pipeline. Execute the agricultural knowledge processing pipeline to obtain an executable work plan.
Citation Information
Cited By
Fragmented semantic understanding order building system and method based on knowledge graph
CN120725026A
Rural socialization service platform based on multi-modal data convergence
CN120849401A
Meteorological intelligent service system based on language large model and construction method
CN121052987A
Plant intelligent factory multi-dimensional data monitoring and collaborative management and control method and system
CN121073151A
Remote sensing agricultural big data management system based on block chain
CN121303603A
