A human-computer collaborative industry chain cognitive graph construction method and system

Abstract

The present application relates to a kind of man-machine collaborative industry chain cognitive graph construction method and system, belong to knowledge graph construction technical field.The present application constructs " expert defines skeleton-machine does complement-manual quality check " four-stage collaborative link for the problem of existing method being inefficient in relying on expert manual definition or being seriously ill by large language model automatic generation hallucination: definer receives the primary industry skeleton defined by expert and is converted into graph data structure;Generator is expanded based on the skeleton and is generated 2~5 level sub-product tree using large language model;Extractor extracts industry chain relationship triple from multi-source heterogeneous industry text;Clusterer is fused coding to semantic vector and graph embedding vector and generates multi-level concept theme tree;Retriever is recalled candidate relationship by three-way heterogeneous retrieval through knowledge graph accurate query, semantic vector retrieval and large language model zero sample generation;Fusioner carries out multi-source fusion, deduplication and confidence normalization to candidate relationship;Association calculator executes relationship inheritance and confidence normalization;Classifier is classified and processed to temporary node;Outputter is pushed to manual review end to final graph, and expert confirms or modifies after feedback into warehouse, realizes closed loop iteration.The present application realizes the deep cooperation of expert knowledge and machine intelligence, greatly improves the construction efficiency while ensuring the quality of graph, and can be widely applied in the field of industry planning, financial risk control, technology service and industrial internet.

Description

Technical Field

[0001] This invention relates to the field of knowledge graph construction technology, and in particular to a human-machine collaborative method and system for constructing a cognitive graph of an industrial chain. Background Technology

[0002] The industrial chain cognitive map is a multi-dimensional semantic network of "technology-industry-economy" with the industry itself as the skeleton, industry entities as nodes, and actual production relations as edges. It is oriented towards the organization and analysis of industrial intelligence and can be widely used in scenarios such as industrial planning research, financial risk control, construction of industrial internet platforms, and transformation of scientific and technological achievements.

[0003] Currently, the construction of industry chain cognitive maps mainly faces the following technical bottlenecks:

[0004] First, the top-down paradigm driven by expert ontology engineering is costly. Existing technologies mostly adopt a top-down expert ontology engineering paradigm, which requires domain experts to predefine the industry chain ontology classification tree. This model is not only costly in terms of manpower, but also difficult to update and expand in a timely manner as new industries continue to emerge and the industry chain structure evolves dynamically, making it difficult to adapt to the needs of rapid industry iteration.

[0005] Secondly, the bottom-up approach driven by large language models suffers from serious illusions. While some recent research has attempted to use large language models for automatic bottom-up knowledge graph construction, these models are prone to fabricating non-existent entities and relationships during the generation process, resulting in unreliable outputs. Although general-purpose large language models possess strong semantic understanding and generalization capabilities, the lack of effective domain knowledge constraints in vertical fields like industry chains makes it difficult to guarantee the accuracy and controllability of extraction results. Furthermore, the knowledge graphs automatically generated by large language models exhibit coarse-grained relationship types and lack hierarchical structure, making it difficult to directly support industry analysis scenarios.

[0006] Third, existing methods fail to effectively integrate expert knowledge and machine intelligence. Purely manual construction is inefficient, while purely automated machine construction lacks quality control. Existing human-machine collaboration solutions either treat humans merely as annotation tools or lack a systematic knowledge feedback loop mechanism. The collaboration between the two remains at a simple division of labor level, failing to achieve deep collaboration of "experts defining the framework—machines completing the completion—humans verifying the quality."

[0007] In view of this, there is an urgent need for a human-machine collaborative technology solution that can fully leverage the advantages of expert knowledge and large language models to achieve the construction of high-quality industry chain cognitive maps. Summary of the Invention

[0008] This invention provides a human-machine collaborative method and system for constructing a cognitive map of the industrial chain, aiming to solve at least some of the problems mentioned above in the prior art.

[0009] To achieve the above objectives, this invention provides a method for constructing a human-machine collaborative industry chain cognitive map, comprising:

[0010] The definer receives the primary industry skeleton defined by the expert and transforms the primary industry skeleton into an initial graph data structure.

[0011] Based on the initial graph data structure, the generator automatically generates a multi-level lower-level product tree using a large language model to obtain a hierarchical skeleton graph.

[0012] A flat knowledge graph is obtained by extracting triplet relationships from multi-source heterogeneous industrial text data using a large language model finely tuned to the vertical domain. The industrial chain relationships include "lower-level product" relationships to express vertical decomposition relationships within an industry and "downstream industry" relationships to express horizontal input-output relationships between industries. The "lower-level product" relationships include six subtypes: sub-products, lower-level terms, functional modules, specifications, application fields, and technology generations. The "downstream industry" relationships include eight subtypes: raw material downstream, production equipment downstream, component downstream, application scenario downstream, process flow downstream, derivative downstream, industrial link downstream, and energy and resource dependent downstream.

[0013] The entities in the flat knowledge graph are encoded by fusing semantic vectors and graph embedding vectors using a clusterer, and a multi-level concept topic tree is generated using a hierarchical clustering algorithm.

[0014] The search engine performs multi-source heterogeneous retrieval on the hierarchical skeleton map and the multi-level concept topic tree to recall candidate industry chain relationship triples.

[0015] The candidate industry chain relationship triples are processed by a fusion engine to perform multi-source fusion, deduplication, and confidence normalization to obtain a complete graph.

[0016] The completed graph is normalized by performing relation inheritance and confidence normalization on the correlation calculator to obtain a normalized graph.

[0017] The temporary nodes in the normalized graph are classified by a classifier to obtain the final industry chain cognitive graph;

[0018] The output device pushes the final industry chain cognitive map to the human review end, and updates the final industry chain cognitive map in response to the confirmation or correction instructions returned by the human review end.

[0019] Preferably, the step of receiving the primary industry framework defined by experts through the definer includes:

[0020] The primary industry knowledge input by experts is transformed into a graph data structure, which includes industry concept entity names, text definition descriptions, and downstream industry information;

[0021] Assign a globally unique identifier to each concept to construct the nodes and edges of the initial knowledge graph.

[0022] Preferably, the step of automatically generating a multi-level subordinate product tree using a large language model via a generator includes:

[0023] Based on a large language model, prompt words are constructed to guide the large language model to generate all subordinate product names of the next level for the current industry concept node;

[0024] Repeat the expansion process for each newly generated child node until a lower-level product tree with four levels, from the second to the fifth level, is generated;

[0025] Establish subordinate product relationships between each newly generated child node and its corresponding parent node.

[0026] Preferably, the multi-source heterogeneous retrieval via the retrieval tool includes:

[0027] First-path retrieval: Using the input keywords, retrieve a preset number of entity or concept nodes with the highest semantic similarity to form a seed set. Use the nodes in the seed set as head entities to query the subordinate and downstream relationships in the hierarchical skeleton graph and the multi-level concept topic tree to obtain tail entities.

[0028] The second retrieval method involves using a query template in parallel retrieval within a vector index constructed from an external corpus to recall the text block most similar to the input keyword.

[0029] The third approach involves calling a large language model to generate candidate relationships based on its own knowledge.

[0030] Preferably, the multi-source fusion, deduplication, and confidence normalization processing via the fusion unit includes:

[0031] The first stage of preliminary merging involves converting all candidate entities into a standard format, merging candidate entities with different names but semantic similarity exceeding a threshold into the same entity, and selecting the entity with the highest confidence score according to the priority order of "precise query of knowledge graph, assisted query of knowledge graph vector, semantic vector retrieval, and generation of large language model".

[0032] The second stage involves fusion of large models: calling large language models to perform semantic duplication merging, classification correction, and text description optimization on the initially merged candidate nodes.

[0033] Preferably, the step of performing relation inheritance and confidence normalization using an association calculator includes:

[0034] Downstream Relationship Inheritance: Using the lower-level product tree constructed by the generator, collect all descendant nodes of the nodes at both ends of the first-level downstream relationship, create a new downstream relationship for each pair of descendant nodes, and decay the confidence level according to the hierarchical depth.

[0035] Confidence normalization: Different confidence calculation methods are used for subordinate product relationships and downstream industry relationships. The supplementary confidence component and the benchmark confidence component are used as inputs, and the comprehensive confidence is calculated by smoothing the product. This ensures that the overall score is not overly penalized when the confidence of a certain source is low, so as to preserve the contribution of multi-source information.

[0036] This invention also provides a human-machine collaborative supply chain cognitive mapping system, comprising:

[0037] The definer module is used to receive the primary industry skeleton defined by experts and convert the primary industry skeleton into an initial graph data structure.

[0038] The generator module is used to automatically generate a multi-level lower-level product tree based on the initial graph data structure using a large language model, thereby obtaining a hierarchical skeleton graph.

[0039] The extractor module is used to extract triplet relationships from multi-source heterogeneous industrial text data using a large language model finely tuned to the vertical domain, resulting in a flat knowledge graph. The industrial chain relationships include "lower-level product" relationships to express vertical decomposition relationships within an industry and "downstream industry" relationships to express horizontal input-output relationships between industries. The "lower-level product" relationships include six sub-types: sub-products, lower-level terms, functional modules, specifications, application fields, and technology generations. The "downstream industry" relationships include eight sub-types: raw material downstream, production equipment downstream, component downstream, application scenario downstream, process flow downstream, derivative downstream, industrial link downstream, and energy and resource dependent downstream.

[0040] The encoder module is used to perform semantic vector encoding on entities and to segment and vectorize the multi-source heterogeneous industrial text data to construct a vector index.

[0041] The clusterer module is used to fuse and encode the entities in the flat knowledge graph with semantic vectors and graph embedding vectors, and to generate a multi-level concept topic tree using a hierarchical clustering algorithm.

[0042] The retrieval module is used to recall candidate industry chain relationship triples through three heterogeneous retrieval methods: precise query from knowledge graph, semantic vector retrieval, and zero-shot generation from a large language model.

[0043] The fusion module is used to perform multi-source fusion, deduplication, and confidence normalization on the candidate industry chain relationship triples;

[0044] The correlation calculator module is used to perform relation inheritance and confidence normalization on the completed graph;

[0045] The classifier module is used to classify temporary nodes in the normalized graph;

[0046] The output module is used to push the final industry chain cognitive map to the human review end, and update the final industry chain cognitive map in response to the confirmation or correction instructions returned by the human review end.

[0047] This invention establishes a four-stage collaborative process: "expert framework definition – machine completion – manual quality inspection," which offers the following advantages compared to existing technologies:

[0048] First, deep human-machine collaboration. This invention receives the first-level industry framework defined by experts through a definer, controlling knowledge quality from the source; it achieves large-scale automated completion through generators, extractors, retrieval units, and fusion units; and it pushes the results to the human review end and receives feedback through an output unit, forming a closed-loop iteration. Experts do not need to build all industry chain relationships line by line; they only need to define the top-level framework and key review nodes, significantly reducing the cost of manual construction.

[0049] Second, enhanced multi-source retrieval. This invention designs a three-way heterogeneous retrieval mechanism—precise query from the knowledge graph, semantic vector retrieval, and zero-shot generation from a large language model—to recall candidate relations from different sources. Through priority-weighted fusion and confidence normalization, it effectively suppresses the illusion problem of the large language model and improves the completion quality.

[0050] Third, semantic-structural fusion clustering. This invention proposes an entity representation method that integrates semantic vectors and graph embedding vectors. It employs a combination strategy of UMAP dimensionality reduction, Hdbscan density clustering, and Ward hierarchical agglomeration for hierarchical clustering, automatically generating multi-level concept topic trees, thus filling the technical gap in the lack of hierarchical organization in industry chain knowledge graphs.

[0051] Fourth, a closed-loop iteration mechanism. Confirmation or correction instructions from the manual review end are fed back to the system for subsequent iterations and optimizations, enabling the quality of the knowledge graph to continuously improve with the number of uses, thus achieving continuous evolution capabilities. Attached Figure Description

[0052] Figure 1 This is a schematic diagram of the overall process of the human-machine collaborative industry chain cognitive map construction method of the present invention.

[0053] Figure 2 This is a schematic diagram of the module structure of the human-machine collaborative industry chain cognitive mapping construction system of the present invention.

[0054] Figure 3 This is a flowchart illustrating the three-way heterogeneous retrieval mechanism of the present invention.

[0055] Figure 4 This is a schematic diagram of the confidence normalization and relation inheritance process of the present invention.

[0056] Figure 5 This is a schematic diagram of the human-machine collaborative closed-loop iterative process of the present invention. Detailed Implementation

[0057] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all of them.

[0058] Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0059] In the description of this invention, the terms "first" and "second" are used only to distinguish different categories of technical features, and not to indicate or imply relative importance or quantity.

[0060] Example 1: System Overall Architecture

[0061] This embodiment details the overall architecture of the human-machine collaborative industry chain cognitive graph construction system of the present invention.

[0062] This invention consists of 10 modules: definer, generator, extractor, encoder, clusterer, retrieval unit, fusion unit, association calculator, classifier, and output unit.

[0063] The knowledge graph in the system undergoes the following transformations sequentially: a first-level industry framework defined by experts, a second to fifth-level lower-level product tree, a set of triples extracted from the corpus, a multi-level concept topic tree generated by clustering, a candidate graph completed through multi-source retrieval, an improved graph with added inheritance relationships and confidence normalization, a graph reclassified from temporary nodes, and finally, the industry chain cognitive graph pushed to the manual review end. All these modules communicate through standardized data interfaces, supporting independent optimization and replacement at the module level.

[0064] Example 2: Definer

[0065] This embodiment details the implementation method of the definer.

[0066] The definer is used to transform the primary industry framework defined by domain experts into a graph data structure that the system can process, and it is the foundation of the entire collaborative construction system.

[0067] The primary industry knowledge input by experts is transformed into a graph data structure, containing industry concept entity names, textual definition descriptions, and downstream industry information. The definer assigns a globally unique identifier to each concept, constructing the nodes and edges of the initial knowledge graph.

[0068] Specifically, the expert input format is an Excel file, with columns defined as the first-level industry name, definition description, and downstream industry name. The definer output is a JSON object; the serialized initial graph contains node and edge information for subsequent module processing.

[0069] Example 3: Generator

[0070] This embodiment details the implementation method of the generator.

[0071] The generator is used to automatically expand and generate a 2 to 5 level subordinate product tree based on an expert-defined primary industry framework.

[0072] Based on a large language model, a prompt-guided model is constructed to generate all subordinate product names at the next level for the current industry concept node. The prompts require each level to generate 5 to 25 child nodes with specific and clear names, and the output format is nested JSON. After generating all subordinate product names at the next level for the current node, the model repeats this process for each newly generated child node until four levels (second to fifth) are generated.

[0073] The node identifier generation rules are as follows: the 2nd and 4th layers (even layers) use numeric suffixes, the 3rd and 5th layers (odd layers) use letter suffixes (A to Z), and if the number of sibling nodes exceeds 26, double-letter encoding is used.

[0074] The newly generated node establishes a "subordinate product" relationship with its parent node. Both the node and the relationship are tagged with the source as "large language model", and the initial confidence level is set to a high value.

[0075] Example 4: Extractor

[0076] This embodiment details the implementation method of the extractor.

[0077] The extractor is used to automatically extract "head entity-relationship-tail entity triples" from unstructured industrial text.

[0078] The relationship type system includes two core relationships: "subordinate products" and "downstream industries". Among them, "subordinate products" are used to express the vertical decomposition structure of the industry from macro concepts to specific products, including the following six subtypes: (1) Sub-products: classification and refinement of products based on materials, formulas, structures or technical routes; (2) Sub-terms: sub-classification of industry concepts based on definitions, attributes or semantic genus relationships; (3) Functional modules: reflecting the system composition relationship from macro whole to local components; (4) Specifications and models: model classification of products based on quantitative parameters such as size, capacity, precision or model; (5) Application fields: field classification of products according to specific uses or industry application scenarios; (6) Technology generations: generational classification of products according to technology evolution routes or version upgrades.

[0079] The term "downstream industry" is used to express the horizontal relationship between industries based on input-output relationships. It includes the following eight subtypes: (1) Raw material downstream: The physical output of industry A is directly used as the production raw material of industry B; (2) Production equipment downstream: The physical output of industry A provides production tools or equipment for industry B; (3) Component downstream: The physical products of industry A are used as components of industry B; (4) Application scenario downstream: The hardware, software or services provided by industry A are applied to the scenarios of industry B; (5) Process flow downstream: Industry A itself is a process or operation step, and its downstream is the next process; (6) Derivative downstream: The output of industry A is transformed into product B through physical or chemical transformation; (7) Industrial link downstream: The overall output of industry A is used as the intermediate input of industry B; (8) Energy resource dependent downstream: Industry A supplies basic energy such as electricity, heat and new energy to industry B.

[0080] The base language model used is Qwen2.5-7B, and LoRA (Low-Rank Adaptation) is employed for vertical domain fine-tuning. The LoRA rank parameter is set to 32, the scaling factor is set to 16, and the dropout is set to 0.05 to 0.1. The adaptation targets are the query projection matrix and the numerical projection matrix of the attention layer in the Transformer architecture.

[0081] The training phase uses the following hyperparameter configuration: 1 training epoch, 8 gradient accumulation steps, and a learning rate of 2.0 × 10⁻⁶. - 4 The optimizer is AdamW, the learning rate is scheduled using a cosine annealing strategy with a warm-up mechanism, and bf16 mixed precision training is used.

[0082] After fine-tuning, the model weights are compressed to a 4-bit fixed-point representation using the GPTQ method. The quantized model can then be deployed on a consumer-grade GPU with 12GB of dedicated video memory.

[0083] For the industry text to be extracted, structured cue words are constructed, including role settings, relation type definitions, and output format constraints. The cue words and the text to be extracted are input into the model, and the model outputs structured triples that conform to the predefined relation type system.

[0084] Example 5: Encoder

[0085] This embodiment details the implementation method of the encoder.

[0086] The encoder is used to map text fragments into low-dimensional continuous vectors to build an efficient index to support near nearest neighbor retrieval.

[0087] For entities, each entity includes three types of textual information: name, definition description, and contextual summary. The encoder repeats the entity name three times and then concatenates it with its definition and summary to form enhanced text. This enhanced text is encoded using a BGE-large-zh-v1.5 pre-trained model to obtain a 768-dimensional entity vector.

[0088] For the external corpus, unstructured documents such as research reports, news, patents, papers, and standards are recursively divided into blocks of 800 characters each with a 150-character overlap to ensure semantic coherence. BGE-large-zh-v1.5 encoding is also used to obtain the corpus block vectors.

[0089] For vector index construction, the FAISS library is used to build the inner product index. All entity vectors and corpus block vectors are merged to build the index. When the total number of vectors exceeds 500,000, IVF inverted index and PQ product quantization are automatically enabled for two-level compression, dividing the vector space into 128 cluster centers. During retrieval, only the 8 clusters closest to the query are accessed. At the same time, each vector is divided into 32 subspaces, and each subspace is independently quantized into 256 codebooks, compressing the storage space to about 1 / 96 of the original while maintaining a recall rate of over 90%.

[0090] Example 6: Clusterer

[0091] This embodiment details the implementation method of the clusterer.

[0092] Clusterers are used to aggregate scattered industry entities into a multi-level conceptual hierarchy tree in a "bottom-up" manner.

[0093] For each entity, its semantic vector and graph embedding vector are concatenated along the feature dimension to form a fused vector. The semantic vector is obtained using BGE-large-zh encoding, and the graph embedding vector is trained using a graph attention network (GAT). The fused vector simultaneously encodes both the entity's textual semantic information and graph structural information.

[0094] Hierarchical clustering employs a three-stage process:

[0095] The first stage is UMAP dimensionality reduction. The UMAP algorithm is used to perform nonlinear dimensionality reduction on the high-dimensional fused vector, while preserving the local neighborhood structure and global topology of the data during the dimensionality reduction process.

[0096] The second stage is Hdbscan density clustering. The Hdbscan algorithm is used to perform cluster analysis on the dimensionality-reduced vector space. It does not require pre-specifying the number of clusters, automatically discovers natural cluster structures in the data and removes noise points, and supports the discovery of clusters of different granularities in different density regions.

[0097] The third stage is Ward hierarchical agglomeration. The Ward linking criterion is used to perform bottom-up hierarchical agglomeration on the clustering results output by Hdbscan, with the goal of minimizing the sum of squared deviations within clusters.

[0098] After clustering is completed, concept nodes are generated for each cluster, lower-level clustering relationships are established between concept nodes and member entities in the cluster, and existing downstream industry relationships are upgraded and merged to generate downstream optimized relationships.

[0099] Example 7: Retrieval Tool

[0100] This embodiment details the implementation method of the search engine.

[0101] The retrieval system is responsible for completing the knowledge of the industry chain and recalling candidate "subordinate products / downstream industries" relationships in parallel through three heterogeneous searches.

[0102] First-line retrieval: Precise knowledge graph query. Using the input keywords, a seed set is created by retrieving a predetermined number of entity or concept nodes with the highest semantic similarity through the FAISS index. Using the seed nodes as head entities, four types of relationships are queried: "subordinate products," "downstream industries," "subordinate clusters," and "downstream optimization," to obtain tail entities. Confidence is calculated based on a combination of the semantic similarity between the head entity and the keywords, and the frequency of the tail entity's occurrence in the corpus.

[0103] The second retrieval method is semantic vector retrieval. A vector index built from an external corpus (research reports, news, patents, etc.) is used for parallel retrieval using query templates to recall multiple text blocks most similar to the query.

[0104] Third-way retrieval: Zero-shot generation using a large language model. This directly calls the large language model to generate candidate relationships based on its own knowledge. A prompt word guides the model to generate possible subordinate products and downstream industries for the keywords, with the output format being a JSON array.

[0105] The results from the three search paths are fed into the fusion processor for further processing.

[0106] Example 8: Fusion Device

[0107] This embodiment details the implementation method of the fusion processor.

[0108] The candidate triplet set output by the retrieval machine suffers from issues such as multi-source duplication, inconsistent entity name representations, and heterogeneous confidence sources. The fusion unit is responsible for merging, deduplicating, and correcting the candidate results, outputting a normalized and complete graph.

[0109] Phase 1: Initial Merging. All candidate entities are converted to a standard format (removing redundant modifiers and standardizing naming styles). Candidate entities with different names but semantic similarity exceeding a threshold are merged into a single entity, retaining the name of the entity with the highest confidence level. When multiple sources point to the same entity, the entity with the highest confidence level is selected according to the priority order: "precise knowledge graph query, knowledge graph vector-assisted query, semantic vector retrieval, and large language model generation."

[0110] Phase Two: Large-Scale Model Fusion. The large language model is invoked to refine the initially merged candidate nodes, performing semantic duplication merging, classification correction, and text description optimization. Cue words guide the model to complete tasks such as merging semantically duplicated nodes, correcting classification errors, optimizing descriptions for greater accuracy and conciseness, and retaining all valuable nodes.

[0111] Example 9: Associated Calculator

[0112] This embodiment details the implementation method of the associated calculator.

[0113] The correlation calculator receives the hierarchical skeleton map output by the generator and the complete map output by the fusion unit, and performs downstream relation inheritance and confidence normalization.

[0114] Regarding downstream relationship inheritance, the first-level downstream relationship set output by the definer contains the associations between first-level industry nodes. The association calculator uses the lower-level product tree constructed by the generator to collect all descendant nodes of levels 2 to 5 for each first-level node, creating a new downstream relationship for each pair of descendant nodes, with the relationship type being the same as the original relationship. The initial confidence of the new relationship decays with the depth of the hierarchy, and the decay is more significant the deeper the hierarchy.

[0115] Regarding confidence normalization, each merged relationship may have multiple sources (expert definition, model generation, retrieval completion, inheritance propagation), with varying confidence scales from each source. This module calculates a comprehensive confidence score for each relationship, uniformly set to the 0-1 range. For subordinate product relationships and downstream industry relationships, different confidence score calculation methods are used, employing retrieval completion components and baseline confidence components as inputs, and calculating the comprehensive confidence score through a smoothed product. When a certain component is low, the overall confidence score is not excessively penalized, preserving the contribution of multi-source information.

[0116] Example 10: Classifier and Output Unit

[0117] This embodiment details the implementation methods of the classifier and the output unit.

[0118] The classifier handles temporary nodes that remain after the association calculator. It categorizes nodes into existing nodes at the same level or groups them into the "Other" category based on their name and semantic similarity. Specifically, it calculates the overall similarity between each temporary node and all nodes at the same level. If the highest similarity reaches a preset threshold, the temporary node is assigned to an existing node; otherwise, if the number of unmatched nodes at a certain level reaches a preset number, a new "Other" category node is created, and all remaining nodes are assigned to it.

[0119] The output device converts the final map output by the classifier into multiple formats for use in different scenarios. Simultaneously, the output device pushes the final map to a human review panel, where industry analysts check, correct, and confirm the automatically generated results through a review interface or the system. Confirmation or correction instructions from the review panel are fed back to the system for subsequent iterative optimization.

[0120] Through the closed-loop mechanism of "experts defining the framework - machine completion - human quality verification", the knowledge graph has achieved continuous iteration and quality improvement.

[0121] Example 11: Application Case

[0122] This embodiment uses the artificial intelligence industry chain as an application scenario to verify the effectiveness of the technical solution of the present invention.

[0123] In terms of data preparation, we obtained multi-source heterogeneous text data in the field of artificial intelligence from the Internet and third-party databases, including industry research reports, portal media news, relevant invention patents, review papers, national standards and encyclopedia entries, totaling more than 50,000 documents.

[0124] In terms of defining the framework, domain experts define the primary framework of the artificial intelligence industry, including primary nodes such as the "AI foundation layer," "AI technology layer," and "AI application layer," as well as their downstream relationships.

[0125] In terms of machine auto-completion, the generator automatically expands the first-level skeleton to generate second to fifth-level subordinate product trees; the extractor extracts more than one million original triples from 50,000 documents; the clusterer generates multi-level concept topic trees; and the retrieval and fusion machines complete candidate relationships through three-way heterogeneous retrieval.

[0126] In terms of the manual review closed loop, the output device pushes the automatically constructed map to the expert review end. Experts check, correct, and confirm the automatically generated results, and the review comments are fed back to the system for iterative optimization.

[0127] Experimental results show that the human-machine collaborative framework of the present invention improves the construction efficiency by about 80% compared with pure manual construction, and reduces the illusion rate of construction quality by about 90% compared with pure automatic construction, thus achieving a dual improvement in efficiency and quality.

[0128] Industrial applicability

[0129] This invention can be applied to the following scenarios: (1) Industrial planning departments can quickly construct industrial chain knowledge graphs to support industrial chain structure analysis and weakness identification; (2) Financial institutions can construct industrial chain risk graphs to conduct industrial chain risk warnings and investment target assessments; (3) Science and technology service platforms can construct achievement-industry matching graphs to achieve precise matching between scientific and technological achievements and industrial needs; (4) Industrial internet platforms can construct supply and demand relationship networks among enterprises. This invention significantly improves construction efficiency while ensuring graph quality through a human-machine collaboration mechanism, and has good prospects for promotion.

[0130] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Any modifications, equivalent substitutions, and improvements 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 constructing a human-machine collaborative industry chain cognitive map, characterized in that, include: The definer receives the primary industry skeleton defined by the expert and transforms the primary industry skeleton into an initial graph data structure. Based on the initial graph data structure, the generator automatically generates a multi-level lower-level product tree using a large language model to obtain a hierarchical skeleton graph. A flat knowledge graph is obtained by extracting triplet relationships from multi-source heterogeneous industrial text data using a large language model finely tuned to the vertical domain. The industrial chain relationships include "lower-level product" relationships to express vertical decomposition relationships within an industry and "downstream industry" relationships to express horizontal input-output relationships between industries. The "lower-level product" relationships include six subtypes: sub-products, lower-level terms, functional modules, specifications, application fields, and technology generations. The "downstream industry" relationships include eight subtypes: raw material downstream, production equipment downstream, component downstream, application scenario downstream, process flow downstream, derivative downstream, industrial link downstream, and energy and resource dependent downstream. The entities in the flat knowledge graph are encoded by fusing semantic vectors and graph embedding vectors using a clusterer, and a multi-level concept topic tree is generated using a hierarchical clustering algorithm. The search engine performs multi-source heterogeneous retrieval on the hierarchical skeleton map and the multi-level concept topic tree to recall candidate industry chain relationship triples. The candidate industry chain relationship triples are processed by a fusion engine to perform multi-source fusion, deduplication, and confidence normalization to obtain a complete graph. The completed graph is subjected to relation inheritance and confidence normalization by an association calculator to obtain a normalized graph. The temporary nodes in the normalized graph are classified by a classifier to obtain the final industry chain cognitive graph; The output device pushes the final industry chain cognitive map to the human review end, and updates the final industry chain cognitive map in response to the confirmation or correction instructions returned by the human review end.

2. The method according to claim 1, characterized in that, The process of receiving the primary industry framework defined by experts through the definer includes: The primary industry knowledge input by experts is transformed into a graph data structure, which includes industry concept entity names, text definition descriptions, and downstream industry information; Assign a globally unique identifier to each concept to construct the nodes and edges of the initial knowledge graph.

3. The method according to claim 1, characterized in that, The automatic generation of a multi-level subordinate product tree using a large language model via a generator includes: Based on a large language model, prompt words are constructed to guide the large language model to generate all subordinate product names of the next level for the current industry concept node; Repeat the expansion process for each newly generated child node until a lower-level product tree with four levels, from the second to the fifth level, is generated; Establish subordinate product relationships between each newly generated child node and its corresponding parent node.

4. The method according to claim 1, characterized in that, The multi-source heterogeneous retrieval via the search engine includes: First-path retrieval: Using the input keywords, a seed set is formed by retrieving a preset number of entities or concept nodes with the highest semantic similarity. Using the nodes in the seed set as head entities, the subordinate and downstream relationships in the hierarchical skeleton graph and the multi-level concept topic tree are queried to obtain tail entities. The second retrieval method involves using a query template in parallel retrieval within a vector index constructed from an external corpus to recall the text block most similar to the input keyword. The third approach involves calling a large language model to generate candidate relationships based on its own knowledge.

5. The method according to claim 1, characterized in that, The multi-source fusion, deduplication, and confidence normalization processes performed by the fusion unit include: The first stage of preliminary merging involves converting all candidate entities into a standard format, merging candidate entities with different names but semantic similarity exceeding a threshold into the same entity, and selecting the entity with the highest confidence score according to the priority order of "precise query of knowledge graph, assisted query of knowledge graph vector, semantic vector retrieval, and generation of large language model". The second stage involves fusion of large models: calling large language models to perform semantic duplication merging, classification correction, and text description optimization on the initially merged candidate nodes.

6. The method according to claim 1, characterized in that, The process of performing relation inheritance and confidence normalization using an association calculator includes: Downstream Relationship Inheritance: Using the lower-level product tree constructed by the generator, collect all descendant nodes of the nodes at both ends of the first-level downstream relationship, create a new downstream relationship for each pair of descendant nodes, and decay the confidence level according to the hierarchical depth. Confidence normalization: Different confidence calculation methods are used for subordinate product relationships and downstream industry relationships. The supplementary confidence component and the benchmark confidence component are used as inputs, and the comprehensive confidence is calculated by smoothing the product. This ensures that the overall score is not overly penalized when the confidence of a certain source is low, so as to preserve the contribution of multi-source information.

7. A human-machine collaborative system for constructing a cognitive graph of an industrial chain, characterized in that, include: The definer module is used to receive the primary industry skeleton defined by experts and convert the primary industry skeleton into an initial graph data structure. The generator module is used to automatically generate a multi-level lower-level product tree based on the initial graph data structure using a large language model, thereby obtaining a hierarchical skeleton graph. The extractor module is used to extract supply chain relationship triples from multi-source heterogeneous industrial text data using a large language model finely tuned to the vertical domain, resulting in a flat knowledge graph. The supply chain relationship includes "lower-level product" relationships to express vertical decomposition relationships within an industry and "downstream industry" relationships to express horizontal input-output relationships between industries. The "lower-level product" relationship includes six subtypes: sub-products, lower-level terms, functional modules, specifications, application fields, and technology generations. The "downstream industry" relationship includes eight subtypes: raw material downstream, production equipment downstream, component downstream, application scenario downstream, process flow downstream, derivative downstream, industrial link downstream, and energy and resource dependent downstream. The encoder module is used to perform semantic vector encoding on entities and to segment and vectorize the multi-source heterogeneous industrial text data to construct a vector index. The clusterer module is used to fuse and encode the entities in the flat knowledge graph with semantic vectors and graph embedding vectors, and to generate a multi-level concept topic tree using a hierarchical clustering algorithm. The retrieval module is used to recall candidate industry chain relationship triples through three heterogeneous retrieval methods: precise query from knowledge graph, semantic vector retrieval, and zero-shot generation from a large language model. The fusion module is used to perform multi-source fusion, deduplication, and confidence normalization on the candidate industry chain relationship triples; The correlation calculator module is used to perform relation inheritance and confidence normalization on the completed graph; The classifier module is used to classify temporary nodes in the normalized graph; The output module is used to push the final industry chain cognitive map to the human review end, and update the final industry chain cognitive map in response to the confirmation or correction instructions returned by the human review end.

8. The system according to claim 7, characterized in that, The retrieval module includes: The knowledge graph query unit is used to retrieve a seed set consisting of a preset number of entities or concept nodes with the highest semantic similarity based on the input keywords, and to query subordinate and downstream relationships using the nodes in the seed set as head entities. A semantic retrieval unit is used to retrieve, in parallel, text blocks that are most similar to the input keywords from the vector index of an external corpus; The large model generation unit is used to call the large language model to generate candidate relations.

9. The system according to claim 7, characterized in that, The fusion module includes: The initial merging unit is used to convert all candidate entities into a standard format, merge candidate entities with different names but semantic similarity exceeding the threshold into the same entity, and select the one with the highest confidence in the priority order of "knowledge graph precise query, knowledge graph vector assisted query, semantic vector retrieval, and large language model generation". The large model refinement unit is used to call the large language model to perform semantic duplication merging, classification correction, and text description optimization on the initially merged candidate nodes.

Images