Unsupervised entity alignment method for highly heterogeneous knowledge graph

Abstract

The invention belongs to the technical field of educational knowledge maps, and discloses an unsupervised entity alignment method for a highly heterogeneous knowledge map, comprising the following steps: constructing a highly heterogeneous knowledge map; the structure-semantic decoupling module is used for extracting and identifying key information and semantic content in the highly heterogeneous knowledge graph; the depth positioning module selects a proper pseudo tag and executes a large-scale artificial intelligence heterogeneous entity alignment task; the knowledge aggregation module integrates and analyzes information obtained by the structure-semantic decoupling module and the depth positioning module, and performs semantic annotation on entities of the data source and relationships between the entities; the structure-semantic decoupling module, the depth positioning module and the knowledge aggregation module cooperate through a self-adaptive region activation mechanism, and the corresponding modules are dynamically activated according to requirements; and outputting the highly heterogeneous knowledge graph after semantic annotation.

Description

Technical Field

[0001] The present invention belongs to the technical field of educational knowledge graphs, and in particular, relates to an unsupervised entity alignment method for highly heterogeneous knowledge graphs. Background Art

[0002] In the field of educational technology, the location search of educational resources (such as textbooks, videos, and exercises) is often based on keywords. However, the location is inaccurate only by keywords. Using knowledge graphs to semantically annotate educational resources allows these resources to be searched based on their content, which can be more accurately located. Entity alignment is a crucial step in semantic annotation.

[0003] Entity alignment (EA) aims to align entities referring to the same real-world objects in different knowledge graphs (KGs). As a key technology to unify heterogeneous data sources, it plays an important role in data-driven artificial intelligence heterogeneous entity alignment (HHEA). Compared with traditional EA, HHEA exhibits significant differences and low overlap in sources, scales (such as entity, relation, and fact counts), and structural features (such as density and structural similarity). In real-world applications, it is often difficult to obtain labeled data (i.e., highly heterogeneous entity pairs) for HHEA tasks, which has prompted research focus to shift to the unsupervised HHEA problem. However, unsupervised HHEA faces significant challenges.

[0004] First, traditional unsupervised EA usually relies on the structural and semantic similarity between KGs to establish alignment. However, in the HHEA scenario, HHKGs show extremely low structural similarity and semantic relevance. The average structural similarity between HHKGs is only 15.4%, and the entity overlap rate is less than 50%. This significant heterogeneity makes it very difficult to obtain reliable structural and semantic relevance between HHKGs.

[0005] Second, in traditional unsupervised EA, due to the low heterogeneity among KGs, specific mechanisms can help establish a clear alignment paradigm to guide the model's decision. However, in HHEA, HHKGs vary greatly in structure and scale, resulting in complex and diverse alignment paradigms. This makes the search space of alignment paradigms extremely large and difficult to evaluate and determine effectively. Therefore, establishing a clear alignment paradigm for HHEA remains a major challenge.

[0006] Currently, the research on higher-order heterogeneous (HHEA) is still in its early stages, and no effective unsupervised HHEA solution has been proposed. Existing HHEA works such as ChatEA adopt a supervised approach. Recently, with the help of the background knowledge and reasoning capabilities of large language models (LLMs), the synergy between LLMs and small models has been achieved. However, simply applying this approach to the unsupervised HHEA scenario cannot effectively solve the above challenges. In addition, schemes like ChatEA combine LLMs and require fine-tuning, which face huge time costs and computing resource pressures in practice. Summary of the invention

[0007] In view of this, this application proposes a new adaptive unsupervised HHEA framework that adopts multi-agent collaboration, combining the semantic understanding ability of large language models (LLMs) and the efficient structural analysis ability of small models. Specifically, inspired by the multi-agent task decomposition paradigm, the framework is decomposed into three functional modules, each of which consists of multiple agent models (LLMs or small models) focusing on different tasks. First, the first functional module focuses on extracting and identifying key information and semantic content in HHKG, providing a basis for subsequent tasks. Next, the second functional module selects appropriate pseudo-labels and performs large-scale HHEA tasks based on the parsed structural and semantic information to optimize the alignment process. Then, the third functional module integrates and analyzes the information obtained in the first two areas, using the expertise of LLMs in background knowledge to supplement the core background knowledge of key entities.

[0008] The three functional modules collaborate through an adaptive region activation mechanism that dynamically activates the corresponding functional modules according to demand. Through the collaborative multi-agent workflow of the three functional modules and their respective agents, the framework effectively integrates the advantages of various models in different tasks and achieves reliable capture of structural and semantic correlations between HHKGs, thereby effectively solving the unsupervised HHEA task.

[0009] In addition, this application proposes a set of optimization tools, including meta-alignment mechanisms and communication protocol rules, to simplify the HHEA process. The meta-alignment mechanism includes meta-expert role-playing instructions and multi-granularity meta-logic symbol rules. The former guides LLMs to dynamically generate appropriate role-playing instructions based on specific highly heterogeneous entities, thereby supplementing the more professional background knowledge of the entities and further constructing reliable structures and semantic correlations between HHKGs. The latter abstracts the complex HHEA paradigm into alignment rules containing semantic and structural information, providing high-quality alignment paradigm guidance for unsupervised HHEA. In addition, this application designs a communication protocol that enhances the interaction efficiency between entities and reduces time and computational costs. Through structured agent interaction and knowledge publish-subscribe mechanism, the interaction efficiency between agents is significantly improved, and time and computational overhead are reduced.

[0010] To achieve the above objectives, the unsupervised entity alignment method of highly heterogeneous knowledge graphs disclosed in this application includes the following steps: Collect educational resources, automatically identify and extract key concepts and entities from educational resources through natural language processing technology, and map them to the defined educational ontology; determine the relationship between entities and represent them in the knowledge graph, thereby constructing a highly heterogeneous knowledge graph; The structure-semantic decoupling module extracts and identifies key information and semantic content in highly heterogeneous knowledge graphs, providing a basis for subsequent tasks; The deep localization module selects appropriate pseudo-labels based on the parsed structural and semantic information and performs large-scale AI heterogeneous entity alignment tasks to optimize the alignment process; The knowledge aggregation module integrates and analyzes the information obtained by the structure-semantic decoupling module and the deep positioning module, uses the large model LLMs to supplement the core background knowledge of key entities, and semantically annotates the entities of the data source and the relationships between entities; The structure-semantic decoupling module, the deep localization module, and the knowledge aggregation module collaborate through an adaptive regional activation mechanism, which dynamically activates the corresponding modules according to demand; Output is a highly heterogeneous knowledge graph after semantic annotation.

[0011] Furthermore, the structure-semantic decoupling module includes three small models: candidate entity retrieval, key node analysis, and extreme information density parsing, which run in a parallel structure to improve the efficiency of the small models in terms of structure and semantic parsing; The deep positioning module selects appropriate pseudo-labels for the key structures and semantic entities parsed, and performs large-scale artificial intelligence heterogeneous entity alignment tasks to obtain a highly heterogeneous knowledge graph similarity matrix; the module includes high-precision heterogeneous entity selection, large-scale precise registration, and multi-granularity induction, which work together in a serial structure; The knowledge aggregation module contains Master and expert prior knowledge, integrates the information of the deep positioning module and the structure-semantic decoupling module, and uses the professional prior knowledge of domain experts to supplement the core background knowledge of key entities, and works collaboratively through a hierarchical structure.

[0012] Furthermore, the adaptive region activation mechanism dynamically adjusts the participation of functional modules by evaluating the current state and activating different functional modules as needed, including structure-semantic decoupling gating, depth positioning gating, and knowledge aggregation gating.

[0013] Furthermore, the structure-semantic decoupling gating determines the activation state of the structure-semantic decoupling module according to the structural difference analysis; specifically, it uses the average entity degree and structural divergence to calculate the difference between two highly heterogeneous knowledge graphs: ; in is the degree difference threshold, is the structural similarity threshold, is the structural similarity score, is the average entity degree of highly heterogeneous knowledge graph 1 and highly heterogeneous knowledge graph 2; when When , the structure-semantic decoupling module enters the inhibition state and is only operated by candidate entity retrieval to maintain basic functions and save computing resources; The depth localization gating controls the activation of the depth localization module by evaluating the input source and its related information; Represents the input source function module, and defines two activation conditions as follows: ; Among them, Area 1 represents the structure-semantic decoupling module, Area 3 represents the knowledge aggregation module, Represents the structure-semantic decoupling module The number of candidate entity sets, of represents the number of new entity pairs with complementary knowledge from the knowledge aggregation module, Represents a logical AND operation; the gate state is determined by the following formula: ; Indicates a logical OR operation. When both conditions are not met, the depth positioning module remains in the inhibited state. , saving computing resources while maintaining system stability; The knowledge aggregation gating determines the activation of the knowledge aggregation module by analyzing the uncertain aligned entity pairs in the similarity matrix: ; in represents the set of uncertainly aligned entity pairs, is the similarity score between entities, Definition i and j The uncertainty interval of an entity, is the minimum threshold of similarity score between entities, is the maximum threshold of similarity scores between entities, is the set of explicitly unaligned entity pairs, Include entities in high-confidence heterogeneous entity alignments; when activated, , unsure alignment of entities is forwarded to the Master for further analysis; otherwise, the module remains suppressed to optimize resource utilization.

[0014] Furthermore, candidate entity retrieval effectively identifies potential candidate entities through similarity-based retrieval, which involves structurally representing the entities and their first-order neighbor information in the highly heterogeneous knowledge graph 2 as text, which is then encoded and stored in a vector database; for each entity in the highly heterogeneous knowledge graph 1, the system retrieves semantically similar entities to form a key candidate entity set that supports subsequent alignment ; The key node analysis converts two highly heterogeneous knowledge graphs into an unweighted directed network, which uses a network key node identification algorithm to locate key entities in the highly heterogeneous knowledge graphs; The extreme information density analysis analyzes entities with extreme information density, and the entities with extreme information density are or Entity, is the minimum threshold of entity node degree, is the maximum threshold of entity node degree, e di is the information density.

[0015] Furthermore, the high-precision heterogeneous entity selection utilizes the semantic understanding capability of a large language model to perform highly heterogeneous entity alignment pairing selection; It processes two types of highly heterogeneous entity inputs: the output from the structure-semantic decoupling module, i.e., the candidate entities of highly heterogeneous entities and their one-hop neighbor information, and the output from the knowledge aggregation module, i.e., the highly heterogeneous knowledge graph and multi-granularity meta-logic symbolic rules after knowledge completion; The large-scale precise alignment utilizes high-confidence heterogeneous entity alignment pairs in high-precision heterogeneous entity selection as training data, and utilizes the efficient computing speed and powerful structure parsing ability of a small model; the model performs large-scale precise entity alignment under supervision; it processes the input highly heterogeneous knowledge graph and aligned entity pairs, generates a comprehensive similarity matrix for all entities, and outputs a similarity matrix between the final aligned entity pairs and the highly heterogeneous knowledge graph; The multi-granularity summary takes as input explicitly aligned entity pairs and explicitly unaligned entity pairs Np to generate multi-granularity meta-logical symbolic rules that capture explicit alignment patterns at different levels of abstraction for subsequent knowledge completion and enhance the understanding of the artificial intelligence heterogeneous entity alignment paradigm in LLM, thereby improving the quality of feedback generated by large models.

[0016] Furthermore, the Master acts as a strategic coordinator of knowledge aggregation, which requires two inputs: uncertain alignment entity pairs Up and multi-granularity meta-logic symbolic rules; the uncertain alignment entity pairs Up are extracted from the similarity matrix, where the similarity scores fall within the range [s1, s2], s1 and s2 are the upper and lower limits of the similarity scores, and are filtered by the knowledge aggregation gating; the Master analyzes these inputs to determine the key knowledge domains required for uncertain alignment and the expert roles required for key knowledge enrichment, and based on this analysis, generates meta-expert role-playing instructions to guide domain experts to complete the task; in addition, the Master assigns specific uncertain alignment entity pairs to the corresponding expert prior knowledge models; The expert prior knowledge model assumes different domain expertise roles; each domain expertise role receives: specific uncertain alignment entity pairs, multi-granularity meta-logic symbol rules, and meta-expert role-playing instructions.

[0017] Furthermore, the multi-granularity meta-logic symbolic rules are used to solve the problem of lack of alignment examples in unsupervised artificial intelligence heterogeneous entity alignment. The rules cover multiple granularity levels and integrate logical reasoning and symbolic representation technology, combining background knowledge and heterogeneous entity information to generate high-confidence heterogeneous entity alignment examples. , as well as explicit misaligned examples; these examples encapsulate the underlying background knowledge required for AI heterogeneous entity alignment, providing critical guidance for subsequent agent tasks; Meta-expert role-playing instructions are used to improve the quality of knowledge augmentation. Specifically, the Master adaptively generates role-playing instructions based on entity types and multi-granular meta-logical symbol rules to guide subsequent LLMs to assume specific expert roles. Furthermore, through structured agent communication interfaces and customized knowledge publish-subscribe mechanisms, communication efficiency is improved and time and computing costs are reduced, including: The shared information pool allows modules to publish and retrieve structured information without directly querying or waiting for responses, significantly improving communication efficiency; To solve the problem of information overload caused by indiscriminate sharing, use a task-priority mechanism: Each module prioritizes information related to its task, extracts relevant data from the shared pool, and stores it in the private pool for task execution; After the task is completed, the module publishes the updated structured data back to the shared pool, enabling efficient collaboration.

[0018] Compared with the prior art, the present invention has the following beneficial effects: This application proposes work that analyzes the challenges of unsupervised highly heterogeneous entity alignment and attempts to address these issues, laying the foundation for future explorations in this complex field; The unsupervised HHEA framework AdaCoAgentEA proposed in this application effectively addresses the challenges of unsupervised HHEA from a multi-agent collaboration perspective. Through the adaptive collaboration of three functional modules formed by multi-agent LLMs and small models, the framework eliminates the dependence on labeled data and effectively captures the structural and semantic correlations between HHKGs; In order to further effectively address the challenges faced by unsupervised HHEA, this application designs a set of optimization tools for AdaCoAgentEA. Specifically, a meta-alignment mechanism is proposed, including meta-expert role-playing instructions and multi-granularity meta-logic symbolic rules, which provides effective information correlation between HHKGs and provides a clear alignment paradigm for unsupervised HHEA, thereby supplementing more professional background knowledge and abstracting complex HHEA scenarios. In addition, this application designs an efficient communication protocol to further improve the efficiency of agent interaction and reduce computational and time costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 The framework of this application,framework structure - semantic decoupling, deep positioning and knowledge aggregation; Figure 2 Detailed implementation process of this application; Figure 3 The communication protocol of this application. DETAILED DESCRIPTION

[0020] The present invention is further described below in conjunction with the accompanying drawings, but the present invention is not limited in any way. Any changes or substitutions made based on the teachings of the present invention belong to the protection scope of the present invention.

[0021] Artificial intelligence is the theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technology in computer science that attempts to understand the essence of intelligence and produce a new intelligent machine that can respond in a similar way to human intelligence. Artificial intelligence is to study the design principles and implementation methods of various intelligent machines so that machines have the functions of perception, reasoning and decision-making.

[0022] Natural language processing (NLP) is an important direction in the fields of computer science and artificial intelligence. It studies various theories and methods that can achieve effective communication between people and computers using natural language. Natural language processing is a science that integrates linguistics, computer science, and mathematics. Therefore, research in this field will involve natural language, that is, the language people use in daily life, so it is closely related to the study of linguistics. Natural language processing technology usually includes text processing, semantic understanding, machine translation, robot question answering, knowledge graph and other technologies.

[0023] The technical solutions provided in the embodiments of the present application involve technologies such as machine learning and natural language processing of artificial intelligence, which are specifically introduced and explained through the following embodiments.

[0024] Before introducing the embodiments of the present application, some terms involved in the present application are first explained.

[0025] 1. Knowledge Graph is a structured representation of real-world knowledge, where Represents a set of entities, represents a set of relations, and Refers to a set of facts. Each fact is Represented as a triple ,in and Is from and Is from .For a temporal knowledge graph, given a set of timestamps , each fact can be associated with a timestamp , thus obtaining a temporal fact representation ,in 2. Entity Alignment (EA): Entity alignment is a basic task in knowledge graph research, which aims to identify corresponding entities in different knowledge graphs. and , the goal is to determine a set of entity pairs , where each pair Represents the same real-world entity in different knowledge graphs.

[0026] like Figure 1 As shown, the framework of this application is decomposed into three key functional modules, namely structure-semantic decoupling, deep positioning and knowledge aggregation: Functional module 1 (structural-semantic decoupling) aims to initially extract and identify key structural and semantic information to provide basic insights for subsequent functional areas. This area includes three small models: candidate entity retrieval (S1), key node analysis (S2), and extreme information density parsing (S3), which run in a parallel structure to maximize the efficiency of small models in structural and semantic parsing; Functional module 2 (deep localization) focuses on selecting appropriate pseudo-labels for the parsed key structural and semantic entities, and performs large-scale HHEA tasks to derive the HHKGs similarity matrix. Under the unsupervised HHEA condition, this process can discover high-quality alignable entities and hidden alignment paradigms. This area includes high-precision heterogeneous entity selection (L1), large-scale precise registration (S4), and multi-granularity induction (L2), which work together in a serial structure; Functional module 3 (knowledge aggregation) integrates the information from the aforementioned fields and uses the professional prior knowledge of Master of Laws (LLMs) and domain experts to supplement the core background knowledge of key entities and work collaboratively through a hierarchical structure.

[0027] These functional modules collaborate according to a program and adopt a demand-based activation mechanism (i.e., a brain-inspired regional adaptive activation mechanism) to evaluate the actual scenario and dynamically activate the necessary functional areas, thereby efficiently coping with the challenges of unsupervised HHEA.

[0028] In one embodiment, the knowledge graph is constructed for educational resources in the following steps: 1. Define educational ontology, including: a. Concept modeling: First, a set of standard educational ontologies needs to be defined, which should include all education-related concepts (such as subject areas, knowledge points, skills, etc.) and the relationships between them. For example, in the field of mathematics, "algebra", "geometry", and "calculus" are different concepts, and there is a hierarchical structure between them. b. Attribute definition: Define a set of attributes for each concept, such as applicable grade, difficulty level, learning time, teaching objectives, etc. These attributes help to describe resources more accurately; 2. Establish unified metadata standards: Metadata framework: Create a unified metadata framework to describe various aspects of educational resources, including but not limited to: Title: the name of the resource. Author / creator: who created the resource. Publication date: when the resource was first published. Update date: when it was last modified or updated. Language: the language used by the resource. Format: the file type of the resource (PDF, video, HTML, etc.). Copyright information: permissions and restrictions on the use of the resource. Semantic tags: Add semantic tags to each resource so that it can be associated with concepts in the knowledge graph. For example, a physics courseware on "Newton's Second Law" might be labeled "Mechanics> Principles of Mechanics> Newton's Laws of Motion"; 3. Constructing knowledge graph: a Entity recognition: Automatically identify and extract key concepts and entities from text materials through natural language processing (NLP) technology, and map them to the defined educational ontology. b Relationship extraction: Determine the relationship between entities and represent them in the knowledge graph. For example, there is a causal relationship between "Newton's second law" and "acceleration" and "force"; c. Linking external data sources: Linking the internally constructed knowledge graph with other publicly available educational resource repositories (such as Wikipedia, academic databases) to enrich content and provide a wider range of learning paths; 4. Labeling and classification: Use existing automatic labeling tools to classify and label educational resources based on the concepts and relationships in the knowledge graph.

[0029] Furthermore, in some embodiments, a series of optimization tools, namely, meta-registration mechanisms and communication protocols, are designed to enhance the robustness of the framework, improve the efficiency of agent interactions, and strengthen the understanding of the registration paradigm under unsupervised HHEA conditions.

[0030] like Figure 2 and Figure 3 As shown, in one embodiment, in order to improve the coordination efficiency between functional areas and reduce the time and computational overhead of unsupervised HHEA, a brain-inspired regional adaptive activation mechanism is introduced, which dynamically adjusts the participation of functional modules by evaluating the current state and activating different functional modules as needed. It is divided into structure-semantic decoupling gating, deep positioning gating, and knowledge aggregation gating.

[0031] Structural-semantic decoupling gating. This gating mechanism determines the activation state of functional module 1 based on structural difference analysis. Specifically, it uses the average entity degree and structural divergence to calculate the difference between two HHKGs. and structural similarity scores : ; in is the degree difference threshold, is the structural similarity threshold. When , the module enters the inhibited state and is operated only by S1 to maintain basic functions while saving computing resources.

[0032] Depth Positioning Gating. This gating mechanism controls the activation of functional module 2 by evaluating the input source and its related information. Representing the input source functional module, we define two activation conditions: ; ; Among them, Area 1 represents the structure-semantic decoupling module, Area 3 represents the knowledge aggregation module, Indicates that from module 1 The number of candidate entity sets, represents the number of new entity pairs with complementary knowledge from module 3. The gated state is determined by the following formula: ; When both conditions are not met, the module remains inhibited , saving computing resources while maintaining system stability.

[0033] Knowledge aggregation gating. This gating mechanism determines the activation of functional module 3 by analyzing the uncertain aligned entity pairs in the similarity matrix: ; in represents a set of uncertainly aligned entity pairs. Here, is the similarity score between entities, The uncertainty interval is defined, is the set of explicitly unaligned entity pairs, Include entities in high confidence heterogeneous entity alignment. When activated When the alignment entity is not determined is forwarded to LLM3 for further analysis; otherwise, the module remains inhibited to optimize resource utilization.

[0034] Agents in structural-semantic decoupling (Functional Module 1). This area contains three specialized agents that collaborate in a parallel structure: Candidate entity retrieval (small model 1) uses the principle that aligned entities usually have similar features to effectively identify potential candidate entities through similarity-based retrieval. This process involves structurally representing the entities and their first-order neighbor information in HHKG2 as text, which is then encoded and stored in a vector database (exemplarily, the database is Faiss, which can be replaced by other databases). For each entity in HHKG1, the system retrieves semantically similar entities to form a key candidate entity set that supports subsequent alignment .

[0035] Critical Node Analysis (Small Model 2) is inspired by complex network analysis and identifies critical nodes in the network to help effectively solve the HHEA task. This agent transforms two HHKGs into unweighted directed networks. It uses network critical node identification algorithms (e.g., betweenness centrality) to quickly locate key entities in the HHKG. This analysis helps effectively solve the HHEA task.

[0036] Extreme Information Density Parsing (Mini-Model 3) specifically analyzes entities with extreme information density to improve the information overload and information deficiency problems encountered during alignment. Two types of entities pose significant challenges to HHEA. First, high-degree entities contain too much information, which may cause information overload and make it difficult to focus on relevant data during alignment. Second, low-degree entities contain insufficient information for accurate alignment judgments. This agent analyzes HHKG to identify entities with extreme information density (calculated as or ), e di It is the information density that facilitates subsequent parsing, thereby improving the efficiency of solving the HHEA task.

[0037] In module 1, agents have the same input and are independent of each other. They can run in a parallel structure, enabling efficient structural and semantic analysis.

[0038] Agents in deep localization (functional module 2). This area contains three specialized agents that collaborate in a serial structure: High-precision Heterogeneous Entity Selection (LLM1) leverages the semantic understanding capabilities of large language models to perform highly heterogeneous entity alignment pairing selection. It can handle two types of highly heterogeneous entity inputs: (1) the output from module 1, i.e., candidate entities of highly heterogeneous entities and their one-hop neighbor information (including nodes with extremely high information density and identified key nodes), and (2) the output from module 3, i.e., HHKG and multi-granularity meta-logic symbolic rules after knowledge completion. Through LLM's powerful contextual reasoning and superior semantic understanding, the agent outputs highly confident heterogeneous entity alignment pairings with superior accuracy.

[0039] Large-scale accurate alignment (small model 4) uses high-confidence heterogeneous entity alignment pairs from high-precision heterogeneous entity selection (LLM1) as training data, and takes advantage of the efficient computational speed and powerful structure parsing capabilities of the small model. The model performs large-scale accurate entity alignment under supervision. It uses Simple-HHEA (one of the HHEA algorithms) to quickly perform large-scale accurate entity alignment within HHKG. It processes the input HHKG and aligned entity pairs to generate a comprehensive similarity matrix for all entities. The output includes the final aligned entity pairs and the similarity matrix between HHKG for subsequent processing.

[0040] Multi-granularity summarization (LLM2), through abstraction and summarization capabilities, analyzes patterns in the HHEA paradigm and generates multi-granularity meta-logic symbolic rules. This agent takes as input the explicitly aligned entity pairs (positive) and explicitly unaligned entity pairs (negative) Np to generate multi-granularity meta-logic symbolic rules. These rules capture explicit alignment patterns at different levels of abstraction, which can be used for subsequent knowledge completion and enhance the understanding of the HHEA paradigm in LLM, thereby improving the quality of feedback generated by large models. Module 2 implements a serial structure in which the output of each agent directly affects the subsequent processing steps.

[0041] Agents in Knowledge Aggregation (Functional Area 3). This area contains two types of specialized agents in the hierarchy: The Master (LLM 3) acts as a strategic coordinator of knowledge aggregation. It requires two key inputs: uncertain aligned entity pairs Up (extracted from the similarity matrix where the similarity scores fall within the range [s1, s2], filtered by the knowledge aggregation gate) and multi-granular meta-logic symbolic rules. The agent analyzes these inputs to identify the key knowledge domains required for uncertain alignment and to determine the expert roles required for key knowledge enrichment. Based on this analysis, it generates meta-expert role-playing instructions to guide domain experts in completing the task. Furthermore, the agent assigns specific uncertain aligned entity pairs to appropriate domain experts, effectively building a highly heterogeneous knowledge aggregation framework that maximizes the collective wisdom of specialized agents while maintaining strategic consistency with the overall HHEA goal.

[0042] Expert prior knowledge models (LLM 4, 5, ...) represent a collection of specialized agents, each of which assumes a different domain expertise role assigned by a master agent. Each domain expertise role receives: (1) specific uncertainly aligned entity pairs, (2) multi-granular meta-logical symbolic rules, and (3) meta-expert role-playing instructions. These expert prior knowledge models work within their designated domains and enrich the HHKG with supplementary knowledge that directly satisfies the requirements specified in multiple aspects.

[0043] Granular meta-logic symbolic rules. This collaborative knowledge enhancement enables HHEA to make more informed decisions by providing key information that may not be obvious in the original HHKG. The output is a set of enhanced HHKGs with knowledge completion, including new entity pairs with supplementary knowledge, thus enabling more accurate alignment decisions in subsequent iterations.

[0044] In module 3, the Master coordinates multiple domain experts through meta-expert role-playing instructions, using a hierarchical structure to promote dynamic knowledge aggregation and refinement.

[0045] Cross-agent workflow. By defining roles, operational skills, and functional scope, we can establish a basic workflow. In this work, we follow the human collaboration SOP and adaptive activation theory of brain partitions to enable all LLM and small model agents to collaborate adaptively and work sequentially to solve the unsupervised HHEA task.

[0046] Specific as Figure 2 and Figure 3 ,After obtaining HHKG, the structural semantic decoupling region, the deep ,localization region and the knowledge aggregation region are activated in sequence after evaluation, ,and subsequently the deep localization region and the knowledge aggregation region are reactivated ,through evaluation and alternating until every region in the workflow is ,evaluated as suppressed.

[0047] Finally, AdaCoAgentEA generates high-quality HHEA results. This application provides a detailed schematic diagram in Ada-CoAgentEA ( Figure 3 ).

[0048] In one embodiment, multi-granularity meta-logic symbolic rules and meta-expert role-playing instructions are used as optimization tools to address the challenges faced by unsupervised HHEA. The meta-alignment mechanism includes multi-granularity meta-logic symbolic rules and meta-expert role-playing instructions, as follows: Multi-granularity meta-logic symbolic rules are used to solve the problem of lack of alignment examples in unsupervised HHEA. The rules cover multiple granularity levels (such as instances and ontologies) and integrate logical reasoning and symbolic representation techniques. As shown in the figure (LLM2), the system needs to combine background knowledge and heterogeneous entity information to generate high-confidence heterogeneous entity alignment instances (positive examples) , as well as explicit misaligned instances (counterexamples). These examples encapsulate the underlying background knowledge required for HHEA and provide critical guidance for subsequent agent tasks.

[0049] Meta-expert role-playing instructions are used to improve the quality of knowledge augmentation. Specifically, LLM3 adaptively generates role-playing instructions based on entity types and multi-granular meta-logical symbolic rules to guide subsequent LLMs to assume specific expert roles. For example, LLM3 can guide LLM4 to play the role of a historian to make full use of its background knowledge in the field related to "Mahmoud_Abbas" instead of choosing the role of a music critic or an ordinary person. In one embodiment, a communication protocol is designed to improve the efficiency of agent communication and reduce time and computational costs through a structured agent communication interface and a customized knowledge publish-subscribe mechanism: Although natural language can support information exchange in LLM-based multi-agent systems, it is easy to cause information distortion in complex unsupervised tasks (similar to the "multi-round phone call" game). The structured agent communication interface of this application adopts the agent information exchange method in the structured communication specification AdaCoAgentEA. Each agent follows a specific pattern and format consistent with its role. The agent that generates multi-granular summaries (LLM2) generates multi-granular meta-logical symbolic rules to capture the core elements of HHEA. Unlike ChatEA, we ensure efficient communication of HHEA in unsupervised scenarios through structured protocols rather than conversational interactions.

[0050] Efficient information dissemination is a core challenge in collaborative systems. The traditional one-to-one communication model is prone to inefficient information flow and error accumulation, especially in unsupervised HHEA, where feedback between agents requires efficient use of information. To this end, the customized knowledge publish-subscribe mechanism of this application uses a framework that combines a global information pool with a private information pool to support collaboration between LLM and smaller models. The shared information pool allows agents to publish and retrieve structured information without directly querying or waiting for responses, which significantly improves communication efficiency. However, indiscriminate sharing may lead to information overload, so we designed a task-priority mechanism: Each agent prioritizes information related to its task, extracts relevant data from the shared pool, and stores it in the private pool for task execution.

[0051] After the task is completed, the agent publishes the updated structured data back to the shared pool, thus achieving efficient collaboration. Figure 3 As shown, the high-precision heterogeneous entity selection agent (LLM1) focuses on the knowledge generated by domain experts and avoids processing irrelevant information published by other agents (such as Master).

[0052] As used herein, the word "preferred" is intended to be used as an example, instance, or illustration. Any aspect or design described herein as "preferred" is not necessarily to be construed as being more advantageous than other aspects or designs. On the contrary, the use of the word "preferred" is intended to present concepts in a specific way. The term "or" as used in this application is intended to mean an inclusive "or" rather than an exclusive "or". That is, unless otherwise specified or clear from the context, "X uses A or B" means any one of the naturally included permutations. That is, if X uses A; X uses B; or X uses both A and B, then "X uses A or B" is satisfied in any of the foregoing examples.

[0053] Moreover, although the present disclosure has been shown and described with respect to one or implementations, those skilled in the art will think of equivalent variations and modifications based on the reading and understanding of this specification and the accompanying drawings. The present disclosure includes all such modifications and variations, and is limited only by the scope of the appended claims. In particular, with respect to the various functions performed by the above-mentioned components (such as elements, etc.), the terms used to describe such components are intended to correspond to any component (unless otherwise indicated) that performs the specified function of the component (such as it is functionally equivalent), even if the structure is not equivalent to the disclosed structure of the function in the exemplary implementation of the present disclosure shown herein. In addition, although the specific features of the present disclosure have been disclosed with respect to only one of several implementations, such features can be combined with one or other features of other implementations that may be desired and advantageous for a given or specific application. Moreover, insofar as the terms "including", "having", "containing" or their variations are used in specific embodiments or claims, such terms are intended to be included in a manner similar to the term "comprising".

[0054] The functional units in the embodiments of the present invention may be integrated into a processing module, or each unit may exist physically separately, or multiple or more units may be integrated into one module. The above-mentioned integrated module may be implemented in the form of hardware or in the form of a software functional module. If the integrated module is implemented in the form of a software functional module and sold or used as an independent product, it may also be stored in a computer-readable storage medium. The above-mentioned storage medium may be a read-only memory, a disk or an optical disk, etc. The above-mentioned devices or systems may execute the storage method in the corresponding method embodiment.

[0055] To sum up, the above embodiment is an implementation mode of the present invention, but the implementation mode of the present invention is not limited by the embodiment. Any other changes, modifications, substitutions, combinations, and simplifications that deviate from the spirit and principles of the present invention should be equivalent replacement methods and are included in the protection scope of the present invention.

Claims

1. An unsupervised entity alignment method for highly heterogeneous knowledge graphs, characterized in that: The following steps are involved: Collect educational resources, automatically identify and extract key concepts and entities from educational resources through natural language processing technology, and map them to the defined educational ontology; determine the relationship between entities and represent them in the knowledge graph, thereby constructing a highly heterogeneous knowledge graph; The structure-semantic decoupling module extracts and identifies key information and semantic content in highly heterogeneous knowledge graphs, providing a basis for subsequent tasks; The deep localization module selects appropriate pseudo-labels based on the parsed structural and semantic information and performs large-scale AI heterogeneous entity alignment tasks to optimize the alignment process; The knowledge aggregation module integrates and analyzes the information obtained by the structure-semantic decoupling module and the deep positioning module, uses the large model LLMs to supplement the core background knowledge of key entities, and semantically annotates the entities of the data source and the relationships between entities; The structure-semantic decoupling module, the deep localization module, and the knowledge aggregation module collaborate through an adaptive regional activation mechanism, which dynamically activates the corresponding modules according to demand; Output is a highly heterogeneous knowledge graph after semantic annotation.

2. The unsupervised entity alignment method based on the highly heterogeneous knowledge graph according to claim 1 is characterized in that: The structure-semantic decoupling module includes three small models: candidate entity retrieval, key node analysis and extreme information density parsing, which run in a parallel structure to improve the efficiency of the small models in structural and semantic parsing; The deep positioning module selects appropriate pseudo-labels for the key structures and semantic entities parsed, and performs large-scale artificial intelligence heterogeneous entity alignment tasks to obtain a highly heterogeneous knowledge graph similarity matrix; the module includes high-precision heterogeneous entity selection, large-scale precise registration, and multi-granularity induction, which work together in a serial structure; The knowledge aggregation module contains Master and expert prior knowledge, integrates the information of the deep positioning module and the structure-semantic decoupling module, and uses the professional prior knowledge of domain experts to supplement the core background knowledge of key entities, and works collaboratively through a hierarchical structure.

3. The unsupervised entity alignment method based on the highly heterogeneous knowledge graph according to claim 2 is characterized in that: The adaptive regional activation mechanism dynamically adjusts the participation of functional modules by evaluating the current state and activating different functional modules on demand, including structure-semantic decoupling gating, depth positioning gating, and knowledge aggregation gating.

4. The unsupervised entity alignment method based on the highly heterogeneous knowledge graph according to claim 3 is characterized in that: The structure-semantic decoupling gating determines the activation state of the structure-semantic decoupling module according to the structural difference analysis; specifically, it uses the average entity degree and structural divergence to calculate the difference between two highly heterogeneous knowledge graphs: ； in is the degree difference threshold, is the structural similarity threshold, is the structural similarity score, is the average entity degree of highly heterogeneous knowledge graph 1 and highly heterogeneous knowledge graph 2; when When , the structure-semantic decoupling module enters the inhibition state and is only operated by candidate entity retrieval to maintain basic functions and save computing resources; The depth localization gating controls the activation of the depth localization module by evaluating the input source and its related information; Represents the input source function module, and defines two activation conditions as follows: ； Among them, Area 1 represents the structure-semantic decoupling module, Area 3 represents the knowledge aggregation module, Represents the structure-semantic decoupling module The number of candidate entity sets, of represents the number of new entity pairs with complementary knowledge from the knowledge aggregation module, Represents a logical AND operation; the gate state is determined by the following formula: ； Indicates a logical OR operation. When both conditions are not met, the depth positioning module remains in the inhibited state. , saving computing resources while maintaining system stability; The knowledge aggregation gating determines the activation of the knowledge aggregation module by analyzing the uncertain aligned entity pairs in the similarity matrix: ； in represents the set of uncertainly aligned entity pairs, is the similarity score between entities, Definition i and j The uncertainty interval of an entity, is the minimum threshold of similarity score between entities, is the maximum threshold of similarity scores between entities, is the set of explicitly unaligned entity pairs, Include entities in high-confidence heterogeneous entity alignments; when activated, , unsure alignment of entities is forwarded to the Master for further analysis; otherwise, the module remains suppressed to optimize resource utilization.

5. The unsupervised entity alignment method based on the highly heterogeneous knowledge graph according to claim 4 is characterized in that: Candidate entity retrieval effectively identifies potential candidate entities through similarity-based retrieval, which involves structurally representing the entities and their first-order neighbor information in the highly heterogeneous knowledge graph 2 as text, which is then encoded and stored in a vector database; for each entity in the highly heterogeneous knowledge graph 1, the system retrieves semantically similar entities to form a key candidate entity set that supports subsequent alignment ; The key node analysis converts two highly heterogeneous knowledge graphs into an unweighted directed network, which uses a network key node identification algorithm to locate key entities in the highly heterogeneous knowledge graphs; The extreme information density analysis analyzes entities with extreme information density, and the entities with extreme information density are or Entity, is the minimum threshold of entity node degree, is the maximum threshold of entity node degree, e di is the information density.

6. The unsupervised entity alignment method based on the highly heterogeneous knowledge graph according to claim 5 is characterized in that: The high-precision heterogeneous entity selection utilizes the semantic understanding capability of a large language model to perform highly heterogeneous entity alignment pairing selection; It processes two types of highly heterogeneous entity inputs: the output from the structure-semantic decoupling module, i.e., the candidate entities of highly heterogeneous entities and their one-hop neighbor information, and the output from the knowledge aggregation module, i.e., the highly heterogeneous knowledge graph and multi-granularity meta-logic symbolic rules after knowledge completion; The large-scale accurate alignment utilizes high-confidence heterogeneous entity alignment pairs in high-precision heterogeneous entity selection as training data, and utilizes the efficient computing speed and powerful structure analysis capability of the small model; The model performs large-scale accurate entity alignment under supervision; It processes the input highly heterogeneous knowledge graph and aligned entity pairs, generates a comprehensive similarity matrix for all entities, and the output includes the similarity matrix between the final aligned entity pairs and the highly heterogeneous knowledge graph; The multi-granularity summary takes as input explicitly aligned entity pairs and explicitly unaligned entity pairs Np to generate multi-granularity meta-logical symbolic rules that capture explicit alignment patterns at different levels of abstraction for subsequent knowledge completion and enhance the understanding of the artificial intelligence heterogeneous entity alignment paradigm in LLM, thereby improving the quality of feedback generated by large models.

7. The unsupervised entity alignment method based on the highly heterogeneous knowledge graph according to claim 6 is characterized in that: The Master acts as a strategic coordinator of knowledge aggregation, which requires two inputs: uncertain alignment entity pairs Up and multi-granularity meta-logic symbolic rules; the uncertain alignment entity pairs Up are extracted from the similarity matrix, where the similarity scores fall within the range [s1, s2], s1 and s2 are the upper and lower limits of the similarity scores, and are filtered by the knowledge aggregation gating; the Master analyzes these inputs to determine the key knowledge domains required for uncertain alignment and the expert roles required for key knowledge enrichment, and based on this analysis, generates meta-expert role-playing instructions to guide domain experts to complete the task; in addition, the Master assigns specific uncertain alignment entity pairs to the corresponding expert prior knowledge models; The expert prior knowledge model assumes different domain expertise roles; each domain expertise role receives: specific uncertain alignment entity pairs, multi-granularity meta-logic symbol rules, and meta-expert role-playing instructions.

8. The unsupervised entity alignment method based on the highly heterogeneous knowledge graph of claim 7 is characterized in that: The multi-granularity meta-logic symbolic rules are used to solve the problem of lack of alignment examples in unsupervised artificial intelligence heterogeneous entity alignment. The rules cover multiple granularity levels and integrate logical reasoning and symbolic representation technology, combining background knowledge and heterogeneous entity information to generate high-confidence heterogeneous entity alignment examples. , as well as explicit misaligned examples; these examples encapsulate the underlying background knowledge required for AI heterogeneous entity alignment, providing critical guidance for subsequent agent tasks; Meta-expert role-playing instructions are used to improve the quality of knowledge augmentation. Specifically, the Master adaptively generates role-playing instructions based on entity types and multi-granularity meta-logical symbol rules to guide subsequent LLMs to assume specific expert roles.

9. The unsupervised entity alignment method based on the highly heterogeneous knowledge graph of claim 8, characterized in that: Improve communication efficiency and reduce time and computing costs through structured agent communication interfaces and customized knowledge publish-subscribe mechanisms, including: The shared information pool allows modules to publish and retrieve structured information without directly querying or waiting for responses, significantly improving communication efficiency; To solve the problem of information overload caused by indiscriminate sharing, use a task-priority mechanism: Each module prioritizes information related to its task, extracts relevant data from the shared pool, and stores it in the private pool for task execution; After the task is completed, the module publishes the updated structured data back to the shared pool, enabling efficient collaboration.

Images