Multi-agent-based bionic domain knowledge graph automatic construction method and system

Abstract

The invention relates to a multi-agent-based bionic domain knowledge graph automatic construction method and system. The method comprises the following steps: constructing a bionic domain conceptual model; establishing a multi-agent system, and automatically extracting bionic entities and semantic relationships between the entities from an unstructured text; carrying out multi-dimensional collaborative verification on the extracted knowledge; based on a Delphi method consultation mechanism, collaborative decision making is carried out on verification results; and performing alignment and fusion on the verified knowledge to form a structured bionic domain knowledge graph. According to the method, three core technologies of a field fine-tuning multi-agent system, a multi-dimensional collaborative verification mechanism and Delphi method consultation decision are fused, so that automatic and high-efficiency construction of the bionic field knowledge graph is realized, and the generated knowledge graph has accuracy and comprehensiveness; the technical problems that a traditional construction method is low in automation degree and the knowledge accuracy is difficult to guarantee are effectively solved, and key knowledge infrastructure support can be provided for bionic theoretical research and technical innovation.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and knowledge engineering technology, specifically to a method and system for automatically constructing a biomimetic domain knowledge graph based on multiple agents. Background Technology

[0002] Biomimetics is an interdisciplinary field that explores the structure, function, and principles of biological systems in nature, drawing inspiration from them to solve engineering and technological challenges. It is a crucial source of disruptive technological innovation. From airplanes invented by mimicking the flight principles of birds to self-cleaning materials developed by drawing inspiration from the hydrophobic properties of lotus leaves, countless inventions stem from profound insights into and imitation of nature. The core of biomimetics research lies in the discovery, transmission, and application of knowledge. However, the inherent complexity of biomimetics knowledge makes it difficult to manage and utilize effectively.

[0003] To effectively manage and utilize this complex biomimetic knowledge, knowledge graph technology offers a promising solution. A knowledge graph, in the form of a graph, uses real-world entities as nodes and relationships between entities as edges, forming a vast semantic network. It can clearly reveal the intrinsic connections between knowledge points, supporting complex knowledge queries, reasoning, and analysis, thereby accelerating the innovation process. However, constructing a high-quality, large-scale biomimetic domain knowledge graph faces significant challenges. Traditional knowledge graph construction methods primarily rely on domain experts for manual reading, annotation, and review. This process not only requires intensive labor input, consuming significant time and human resources, but is also inefficient and struggles to keep pace with the explosive growth of knowledge. Furthermore, the manual construction process is prone to introducing subjective biases and inconsistencies, resulting in inconsistent knowledge graph quality and poor scalability.

[0004] In recent years, the rapid development of natural language processing technologies, represented by large language models (LLMs), has brought new hope to the construction of automated knowledge graphs. LLMs have demonstrated powerful capabilities in text understanding, information extraction, and content generation. Researchers have begun to explore using LLMs to directly extract entities and relationships from unstructured text, hoping to replace the traditional manual annotation process.

[0005] Existing technologies include several attempts to leverage artificial intelligence to assist in biomimetic knowledge management or design. For example, Chinese patent document CN117057173B discloses a biomimetic design method supporting divergent thinking. This method uses the GPT-3 model to extract "source, benefit, application" triples and constructs a semantic network, achieving biomimetic knowledge retrieval and mapping by calculating semantic similarity. However, this method primarily focuses on retrieving design inspiration and divergent thinking, lacking a rigorous fact-checking mechanism for model-generated content (especially hallucination phenomena), and its mapping mainly relies on vector similarity, resulting in a relatively singular decision-making dimension. Another example is Chinese patent document CN114610846A, which discloses a knowledge graph expansion and completion method based on a heuristic biomimetic knowledge grafting strategy. This method utilizes SAO grammatical structure analysis and the K-BERT model for knowledge injection and expansion. However, this method highly relies on fixed grammatical structures (such as subject-verb-object SAO), lacks flexibility in handling complex semantics, and lacks dynamic collaboration and consultation capabilities among multiple agents, making it difficult to automatically correct errors during the construction process.

[0006] In summary, despite significant progress in existing research, directly applying general LLMs to the construction of specialized biomimetic knowledge graphs still faces the following challenges:

[0007] (1) Insufficient domain adaptability and specialization: General LLMs lack in-depth knowledge of the specific field of bionics. When dealing with highly specialized terminology, complex biological mechanisms and interdisciplinary connections, their accuracy in understanding and extracting information will decrease significantly, which may lead to confusion of concepts or failure to identify implicit connections.

[0008] (2) Knowledge Accuracy and Hallucination Issues: When generating content, LLMs sometimes fabricate information that does not conform to the facts, which is called hallucination, and existing technologies often ignore this problem. In the task of knowledge graph construction, which has extremely high requirements for factual accuracy, the hallucination problem is fatal. A graph full of false knowledge can even mislead scientific research and innovation activities.

[0009] (3) Lack of effective verification and correction mechanisms: Most existing methods focus on how to extract information from text, but lack mature and reliable mechanisms for systematically verifying the accuracy, consistency and authenticity of the extracted information. Simple post-processing rules are difficult to cope with the complexity of knowledge and the diversity of LLM illusions.

[0010] (4) The decision-making mechanism is simple and lacks flexibility: Existing methods (such as Chinese patent document CN 114610846A) usually use simple weighted summation, similarity matching or hard threshold to synthesize the results. This mechanism cannot fully consider the mutual influence and dynamic adjustment between various verification dimensions, and is difficult to cope with complex and ever-changing knowledge verification scenarios.

[0011] (5) Inconsistent knowledge representation and chaotic structure: Without the guidance of a predefined and unified knowledge framework (i.e., conceptual model or schema), entities and relations generated by different LLMs or different extraction batches may differ in naming, granularity and classification, resulting in a chaotic structure and inconsistent semantics in the final constructed knowledge graph, making it difficult to effectively integrate and compute.

[0012] Therefore, there is an urgent need to propose a new automated knowledge graph construction method that can fully utilize the powerful capabilities of LLMs while ensuring domain adaptability, knowledge accuracy, and representation consistency, and integrate multi-dimensional verification results through a more flexible and intelligent decision-making mechanism to meet the actual needs of the development of the bionics field. Summary of the Invention

[0013] The purpose of this invention is to provide a method and system for automatically constructing a biomimetic domain knowledge graph based on multi-agent technology. This method and system can solve the core problems existing in the prior art, such as low automation, high labor costs, difficulty in ensuring knowledge accuracy, poor domain adaptability, single decision-making mechanism, and inconsistent knowledge representation. It aims to efficiently and accurately construct a biomimetic domain knowledge graph with clear structure, reliable content, and broad coverage from large-scale, unstructured text data.

[0014] To achieve the above objectives, the technical solution of the present invention is as follows:

[0015] A method for automatically constructing a biomimetic domain knowledge graph based on multi-agent systems, comprising the following steps:

[0016] S1. Construct a conceptual model for the field of bionics;

[0017] S2. Based on the bionic domain conceptual model, construct a multi-agent system, and use the multi-agent system to automatically extract the semantic relationships between bionic entities from unstructured text, and output the entity set and relation triplet that conform to the model specification.

[0018] S3. Utilize the multi-agent system to perform multi-dimensional collaborative verification of the biomimetic entity and the semantic relationships between entities;

[0019] S4. Based on the Delphi method consultation mechanism, collaborative decision-making is carried out on the verification results. After multiple rounds of anonymous feedback and opinion revision, a consensus is reached, and a final judgment is made on the validity of the knowledge.

[0020] S5. Align and merge the verified entities and relationships, and store them in a graph database to form a structured biomimetic domain knowledge graph.

[0021] As a further improvement to the above technical solution, step S1 specifically includes:

[0022] S11. Define the core entity type;

[0023] The conceptual model in the biomimetic domain predefines core entity types and relation types. The core entity types include: BioEntity, BioComponent, BioStrategy, Function, InnovationPrinciple, Innovation, Challenge, and ApplicationPotential. This conceptual model can be represented as a graph structure O=(C,R), where C is the set of entity types and R is the set of relation types.

[0024] S12, Define the relation type;

[0025] The relationship types are used to define various semantic relationships between entities; for example, belongToBio, sourceFrom, haveFunction, generateInnovation, etc., which together constitute a rigorous semantic network framework.

[0026] As a further improvement to the above technical solution, step S2 specifically includes:

[0027] S21. Construct a multi-agent system; the multi-agent system includes an entity extraction agent and a relation extraction agent, wherein each agent is obtained by fine-tuning a pre-trained large language model through low-rank adaptive (LoRA) technology to possess professional knowledge in the field of bionics.

[0028] S22. Perform entity extraction; the entity extraction agent is responsible for identifying and extracting the eight entity types defined in step S11 from the input text. This extraction process can be represented as follows:

[0029] ;

[0030] Where E represents the extracted entity set, x is the input free text, O represents the biomimetic domain concept model, and LLM... LoRA This represents the large model after LoRA fine-tuning.

[0031] S23. Perform relation extraction: The relation extraction agent first selects a set of candidate entity pairs P that meet the relation constraints based on the relation types defined in the biomimetic domain conceptual model. This process first generates a set of candidate entity pairs P based on the entity type constraints defined in the conceptual model, then performs relation judgment on the candidate entity pairs, and outputs the relation triplet to be verified.

[0032] The set of candidate entity pairs P is shown in the following equation:

[0033] ;

[0034] in, and Representing two entities, E type1 E type2 represents the entity type in the conceptual model, r represents the specific relationship type between entity types defined in the conceptual model, and O(R) represents the set of relationships defined in the schema.

[0035] As a further improvement to the above technical solution, step S3 specifically includes:

[0036] Step S3 specifically includes:

[0037] S31. Perform context consistency verification on the biomimetic entities and semantic relationships extracted in step S2: Calculate the semantic similarity between the knowledge to be verified and the original unstructured text context using a semantic similarity calculation model, determine the correlation between the two, and output the context consistency verification result.

[0038] For entity e i Context consistency verification is performed using the following formula:

[0039] ;

[0040] in, This indicates the entity context consistency verification result. The context text representing the source of entity extraction. This represents a context-verification agent based on a large language model.

[0041] For relation triples (e i e j ,r), use the following formula to perform context consistency verification:

[0042] ;

[0043] in, and They represent two entities respectively. Indicates a relationship.

[0044] S32. Utilize a multi-agent system to verify the consistency of the conceptual model between the biomimetic entities and their semantic relationships extracted in step S2: This is done by checking whether the extracted entity types and relationship types conform to the predefined constraints in the biomimetic domain conceptual model. This process can be represented as:

[0045] ;

[0046] in, This indicates the entity schema consistency verification result. For a schema verification agent based on a large language model, v schema This represents the entity schema consistency verification result, and O represents the biomimetic conceptual model.

[0047] S33. Use a multi-agent system to verify the factual consistency of the biomimetic entities and semantic relationships extracted in step S2: By calling the external network search tool Tool, the extracted knowledge is fact-checked and confirmed.

[0048] For entities The following formula is used to verify the consistency of facts:

[0049] ;

[0050] in, This indicates the result of the entity fact consistency verification. This refers to the external network search tools invoked by the intelligent agent. A fact-verifying agent based on a large language model, equipped with web search tools.

[0051] For relation triples (e i e j The following formula is used to verify the consistency of facts:

[0052] ;

[0053] in, LLM represents the result of the entity relationship factual consistency verification. fact A fact-verifying agent based on a large language model, equipped with web search tools.

[0054] As a further improvement to the above technical solution, step S4 specifically includes:

[0055] S41. Initialize expert agent group: Set up multiple parallel decision agents of conservative, balanced and intelligent types. Each decision agent receives all the verification results of step S3 and makes an independent comprehensive judgment based on its own preset decision strategy, and outputs a preliminary decision score and decision reason; each expert agent gives a preliminary judgment independently based on different verification dimensions or domain knowledge background.

[0056] S42. Multi-round anonymous consultation and iteration: The Delphi method is applied to the preliminary decision results of multiple decision agents for consultation. Through multiple rounds of anonymous feedback and correction judgment, a consensus is gradually reached. In each round of consultation, each decision agent receives the anonymous decision results and reasons of other agents, and adjusts its own judgment according to the group opinion, dynamically adjusting its own confidence and judgment conclusion, and completing multiple rounds of correction and iteration.

[0057] S43. Consistency Judgment and Iteration Termination: After each round of consultation, the consistency index of expert opinions is calculated. If the consistency reaches the preset threshold, or the number of iterations reaches the maximum preset value, the iteration is terminated.

[0058] S44. Final Decision Formation: A final decision is formed based on the consultation results. If the opinions of the decision-making agents converge, the final validity score is formed by aggregating the expert judgment results of the final round through a voting mechanism or weighted average method. If the opinions of the decision-making agents do not converge, the knowledge is marked as pending manual review. At the same time, a detailed report containing the decision-making process of each round and the core points of disagreement is automatically generated to provide a basis for manual review. The final determination of the validity of the knowledge is completed in sequence.

[0059] As a further improvement to the above technical solution, in step S42, the judgment and adjustment process of the expert agent adopts a feedback-based dynamic adjustment mechanism, in which each expert agent updates its confidence level according to the opinions of other experts.

[0060] As a further improvement to the above technical solution, in step S43, the consistency index is measured using the coefficient of variation (CV).

[0061] As a further improvement to the above technical solution, step S5 specifically includes:

[0062] S51. Knowledge Alignment and Fusion: Using a semantic vector-based approach, all extracted entities are encoded into high-dimensional semantic vectors through a language model. By calculating the cosine similarity between vectors, entities pointing to the same concept are automatically identified and merged to eliminate redundancy.

[0063] The formula for calculating the cosine similarity is:

[0064] ;

[0065] Where Sim represents the similarity between two vectors, and V i and V j Let represent the semantic vectors of the two entities respectively, · represent the vector dot product, and ||·|| represent the L2 norm of the vector. The extracted entities are encoded into semantic vectors, and the cosine similarity between the vectors is calculated to identify and merge duplicate or equivalent knowledge entries in order to eliminate knowledge redundancy.

[0066] S52. Knowledge Storage: The aligned and fused knowledge is loaded into a graph database for persistent storage. Entities are stored as nodes in the graph database, and the semantic relationships between entities are stored as edges between nodes, ultimately forming a structured biomimetic domain knowledge graph.

[0067] As a further improvement to the above technical solution, the graph database is a Neo4j graph database, which is used to store and manage entities, relationships and their attributes, and supports efficient knowledge query and reasoning.

[0068] In a second aspect of the invention, a biomimetic domain knowledge graph automatic construction system based on multi-agent systems is disclosed, the system comprising:

[0069] The concept model building module is used to build concept models in the field of bionics.

[0070] The knowledge extraction module is used to construct a multi-agent system based on the bionic domain concept model, and use the multi-agent system to automatically extract the semantic relationships between bionic entities from unstructured text, and output the entity set and relation triplet that conform to the model specification.

[0071] The knowledge verification module is used to perform multi-dimensional collaborative verification of the biomimetic entity and the semantic relationships between entities using the multi-agent system;

[0072] The Delphi method consultation and decision-making module is used to make collaborative decisions on the verification results based on the Delphi method consultation mechanism. After multiple rounds of anonymous feedback and opinion revision, a consensus is reached, and a final judgment is made on the validity of the knowledge.

[0073] The knowledge graph generation module is used to align and merge verified entities and relationships, store them in a graph database, and form a structured biomimetic domain knowledge graph.

[0074] Compared with the prior art, the advantages of the present invention are:

[0075] (1) This invention proposes a biomimetic domain knowledge graph automatic construction method based on multi-agent system. By adopting a multi-agent system based on fine-tuning LLM, the knowledge extraction, verification and integration process that previously required a lot of manual participation is fully automated, which greatly shortens the knowledge graph construction cycle and compresses the work of several months or even several years into several hours or several days, significantly improving the construction efficiency.

[0076] (2) This invention designs an innovative multi-dimensional collaborative verification mechanism, which provides triple protection for knowledge quality from three aspects: fidelity to the original text, conformity to domain logic, and consistency with external facts. This mechanism can effectively filter out the "illusion" and extraction errors generated by LLM, ensuring the high accuracy and reliability of the final knowledge stored in the database, and solving the core pain point of the difficulty in guaranteeing the accuracy of knowledge in the prior art.

[0077] (3) This invention achieves deep adaptation to the highly specialized field of bionics by constructing a domain-specific conceptual model and fine-tuning the LLM. This method can accurately understand domain terminology and capture complex professional logic. Its performance far exceeds that of an unmodified general-purpose large language model, ensuring the professionalism and depth of the knowledge graph.

[0078] (4) This invention introduces a consultation and decision-making mechanism based on the Delphi method. Through multiple rounds of expert opinion solicitation and feedback, each verification agent can learn from each other and make dynamic adjustments, avoiding the limitations of traditional hard weighted decision-making, significantly improving the flexibility and accuracy of decision-making, and enabling the system to better cope with complex and ever-changing knowledge verification scenarios.

[0079] (5) The knowledge graph constructed in this invention is highly consistent in structure, semantically clear, and free of redundancy. Since all extraction work is carried out under the guidance of a unified conceptual model and has undergone strict alignment and fusion, it lays a solid foundation for subsequent knowledge query, complex reasoning, and data mining applications, thereby improving the usability of the knowledge graph. Attached Figure Description

[0080] Figure 1 This is a flowchart of the method for automatically constructing a biomimetic domain knowledge graph based on multiple agents in this invention;

[0081] Figure 2 This is a schematic diagram of the overall process of the biomimetic domain knowledge graph automatic construction method based on multi-agent in this invention;

[0082] Figure 3 This is a schematic diagram of entity relationships in the conceptual model (Schema) of the biomimetic field in this invention;

[0083] Figure 4 This is a schematic diagram of the multi-dimensional collaborative verification mechanism in this invention;

[0084] Figure 5 This is a flowchart illustrating the Delphi method consultation and decision-making mechanism in this invention;

[0085] Figure 6 This is a schematic diagram of the entity alignment process in this invention. Detailed Implementation

[0086] To provide a better understanding of the structural features and effects achieved by the present invention, a detailed description is provided below, accompanied by preferred embodiments and accompanying drawings:

[0087] This invention relates to the fields of artificial intelligence and knowledge engineering, specifically to a method and system for automatically constructing a knowledge graph in the field of bionics based on a multi-agent system and a large language model. In particular, this invention relates to how to efficiently and accurately extract bionics-related entities, relationships, and knowledge from massive, unstructured, multi-source heterogeneous text data, and structure them into a high-quality knowledge graph to serve applications such as bionics research, technological innovation, and interdisciplinary education.

[0088] Example 1

[0089] This embodiment provides a method for automatically constructing a biomimetic domain knowledge graph based on multiple agents, such as... Figure 1 and Figure 2 As shown, the method includes core steps such as data preprocessing, knowledge extraction, knowledge verification, Delphi method consultation and decision-making, knowledge fusion and storage.

[0090] S0, Data Preprocessing Module

[0091] The data preprocessing module is the starting point of the entire knowledge graph construction process, responsible for collecting and processing the raw data used for knowledge extraction. In this embodiment, the data source is mainly a massive amount of unstructured text related to bionics, such as academic journal articles, patent specifications, science and technology news, and biological encyclopedias.

[0092] S01, Data Collection Strategy

[0093] This embodiment employs a multi-source heterogeneous data collection strategy to ensure the comprehensiveness and authority of the knowledge graph. Specifically, the system automatically retrieves relevant documents from multiple pre-defined high-quality information sources through web crawlers, API interfaces, and other methods. The list of information sources includes, but is not limited to:

[0094] (1) Academic journal databases: PubMed (biomedical literature), IEEE Xplore (engineering and technology literature), ScienceDirect (multidisciplinary literature), etc. These databases provide high-quality peer-reviewed papers and are the core source of bionics knowledge.

[0095] (2) Patent databases: USPTO (United States Patent and Trademark Office), EPO (European Patent Office), CNKI patent database, etc. Patent documents record in detail the specific applications and innovative solutions of biomimetic technology, and are an important bridge connecting bio-inspiration and engineering implementation.

[0096] (3) Encyclopedia knowledge bases: Wikipedia, Encyclopedia of Life, Baidu Baike, etc. These resources provide basic descriptions and taxonomic information of biological entities.

[0097] (4) Science and technology news and blogs: Nature News, Science Daily, MIT Technology Review, etc. These channels report the latest research progress and technological breakthroughs in bionics.

[0098] To improve the accuracy of data collection, the system is configured with a keyword list, including core terms such as "biomimetics," "bio-inspired," "nature-inspired," and "biomimicry," as well as specific biological names (such as "gecko," "lotus leaf," and "shark skin"). When crawling, the web crawler prioritizes documents whose titles or abstracts contain these keywords.

[0099] S02, Text Cleaning and Normalization

[0100] The original webpage or PDF document contains a large amount of non-text information, such as HTML tags, CSS styles, JavaScript scripts, images, table formatting code, etc. This step uses a professional parsing tool library to remove this irrelevant metadata, retaining only the plain text content.

[0101] (1) PDF document parsing: For PDF files, this embodiment uses the PyMuPDF (also known as Fitz) library. This library can not only efficiently extract text, but also extract table content and convert it into structured Markdown format.

[0102] (2) HTML document parsing: For HTML format web pages, this embodiment uses the lxml library for high-speed parsing. The system extracts the text in the main content tags first based on the DOM tree structure, while filtering out navigation, advertisements, footers and script areas.

[0103] (3) Text repair and standardization: The system uses a series of regular expressions and heuristic rules to handle common problems. For example, it automatically repairs word breaks caused by line breaks in PDFs (such as "know-ledge" → "knowledge"), unifies various citation formats into a standard format, removes extra spaces and line breaks, and converts all text to UTF-8 encoding.

[0104] S03, Semantic Blocking Strategy

[0105] Directly inputting long documents into an LLM for processing is not only inefficient but may also exceed the model's context length limit (e.g., 8K tokens). Therefore, it is necessary to segment long texts into smaller units with complete semantic meaning. This embodiment employs an advanced hierarchical semantic chunking strategy.

[0106] (1) Macro-structure segmentation: First, the system uses the document's structured tags (such as chapter titles and paragraph tags) for preliminary segmentation. For academic papers, the system will recognize standard chapters such as "Abstract", "Introduction", "Methods", "Results", and "Discussion".

[0107] (2) Micro-level semantic boundary detection: Within each chapter or paragraph, the system applies a semantic boundary detection algorithm based on sentence embedding. Specifically, the all-mpnet-base-v2 model from the sentence-transformers library is used, which performs well on semantic similarity tasks. The system first segments the paragraph into sentences, and then generates a 768-dimensional semantic vector for each sentence. Next, the cosine similarity of each pair of adjacent sentence vectors is calculated to form a similarity curve. To smooth out noise, the system uses a moving average filter with a window size of 3. Finally, the local minimum points of the similarity curve are found through a peak detection algorithm. These points correspond to the transformation boundaries of semantic topics, and the system performs segmentation at these points.

[0108] (3) Block size control: In order to achieve the best balance between LLM processing efficiency and context integrity, the system controls the length of each text block to be between 256 and 512 tokens.

[0109] After the above three layers of processing, the system obtains a series of clean, standardized, and semantically complete text blocks, each with metadata such as its source document, chapter information, and position in the original text. These text blocks will serve as input for the subsequent knowledge extraction module.

[0110] S1. Constructing a conceptual model for the field of biomimetics

[0111] In this embodiment, constructing a biomimetic domain conceptual model (Schema) is the top-level design of the entire process, aiming to provide a unified and clear semantic framework for subsequent automated knowledge extraction and knowledge graph construction. This model defines the types and attributes of knowledge, as well as the legitimate relationships between them, and is the cornerstone for ensuring the structural consistency and semantic accuracy of the knowledge graph.

[0112] S11. Define the core entity type

[0113] Based on in-depth analysis of literature in the field of bionics and knowledge review by experts in the field, this embodiment defines eight core entity types, covering the complete knowledge chain from bio-inspiration to technological innovation. The specific definitions are shown in the table below:

[0114]

[0115] S12, Define relation types

[0116] To describe the relationships between entities, this embodiment defines a series of semantic relationship types, such as Figure 3 As shown in the table below, each relationship explicitly defines the legal entity types of its subject (Domain) and object (Range), forming strict logical constraints. The specific definitions are as follows:

[0117]

[0118] S2. Knowledge Extraction Based on Multi-Agent Systems

[0119] This step is responsible for automatically extracting knowledge that conforms to the conceptual model defined in S1 from the preprocessed text. Its core is to build a multi-agent system (MAG) consisting of multiple specialized agents.

[0120] S21. Constructing a multi-agent system

[0121] The agent in this embodiment is built upon a powerful pre-trained large language model (such as Qwen2.5-32B-Instruct). To imbue it with expertise in the field of bionics, this invention employs a parameter-efficient low-rank adaptive (LoRA) technique for domain fine-tuning.

[0122] LoRA (Local Area Replication) technology efficiently injects new knowledge into the model while freezing the original model parameters by adding trainable low-rank matrices (A and B) alongside the model's existing weight matrix. For a pre-trained weight matrix W0 (with dimensions d×k), LoRA does not directly update... Instead, a "bypass" is added next to it, using the product of two low-rank matrices A (dimension d×r) and B (dimension r×k) to approximate the weight update amount. , that is:

[0123] ;

[0124] Among them, the rank r is a hyperparameter much smaller than d and k (r << min(d, k)). During training, only the parameters of A and B are updated, while W0 remains frozen. In this way, the number of parameters to be trained is reduced from d×k to r*(d + k), greatly reducing the training cost and hardware requirements.

[0125] During the forward propagation process, the output of the model is calculated as:

[0126] ;

[0127] Among them, x is the input vector and h is the output vector.

[0128] To control the influence intensity of the LoRA layer, a scaling factor α is usually introduced, such that:

[0129] ;

[0130] In this embodiment, the configuration of LoRA is as follows:

[0131] (1) Rank (r): Set to 16. This is an empirical value that achieves a good balance between performance and efficiency.

[0132] (2) Alpha ( ): The scaling factor of LoRA, set to 32. The ratio of is used to scale

[0133] (3) Target module: Apply LoRA to all linear layers in the Transformer model, including the fully connected layers of query, key, value, and output, to maximize its ability to adjust the model behavior.

[0134] (4) Dropout: Set a dropout rate of 0.05 in the LoRA layer to prevent overfitting.

[0135] The step S21 specifically includes:

[0136] S211. Construction of the domain fine-tuning dataset

[0137] This embodiment constructs a biomimetic instruction fine-tuning dataset containing approximately 10,000 samples. Each sample is a JSON object containing three fields: instruction, input text, and output (expected output). Data sources include:

[0138] (1) Manually written: High-quality instructions and question-and-answer pairs were written by three experts in the field of bionics (with backgrounds in biology, materials science and mechanical engineering, respectively) based on typical scenarios. This part of the data accounts for about 20% of the total dataset, ensuring a high-quality benchmark for the data.

[0139] (2) Semi-automatic generation: Using more powerful models such as GPT-4o, a large amount of synthetic data is generated based on a small number of seed samples (i.e., the manually compiled samples mentioned above). This part of the data accounts for about 60%.

[0140] (3) Data augmentation: Perform operations such as synonym replacement, sentence restructuring, and entity replacement on existing samples to increase data diversity. This part of the data accounts for about 20%.

[0141] The dataset is divided into training, validation, and test sets in an 8:1:1 ratio.

[0142] S212, Training Process and Hyperparameter Settings

[0143] This embodiment uses the Hugging Face transformers library (version 4.30.0) and the peft library (Parameter-Efficient Fine-Tuning, version 0.4.0) for training. Training was performed on a server equipped with eight NVIDIA A100 80GB GPUs, running Ubuntu 22.04 with CUDA version 12.1.

[0144] The training hyperparameters are set as follows:

[0145] (1) Optimizer: AdamW, learning rate 2×10 -4 The weight decays to 0.01.

[0146] (2) Learning rate scheduling: A linear warmup strategy is adopted, with the number of warmup steps being 10% of the total number of training steps, and then linearly decaying to 0.

[0147] (3) Batch Size: The batch size for each GPU is 8. Gradient accumulation is used with a cumulative step of 2. Therefore, the effective batch size is 8×8×2=128.

[0148] (4) Training epochs: 3 epochs.

[0149] (5) Maximum sequence length: 512 tokens.

[0150] (6) Mixed precision training: Use FP16 (16-bit floating-point) mixed precision training to accelerate training and reduce memory usage.

[0151] The entire fine-tuning process took approximately 48 hours. During training, the system evaluated the model performance on the validation set every 500 steps and saved the best-performing checkpoint. Finally, the model with the highest F1 score on the validation set was selected as the deployment model.

[0152] S22, Perform entity extraction

[0153] The entity extraction agent receives text blocks as input and guides the LLM to perform tasks through carefully designed prompts. This extraction process can be represented as:

[0154] ;

[0155] Where E represents the extracted entity set, x is the input free text, O represents the biomimetic domain concept model, and LLM... LoRA This represents the large model after LoRA fine-tuning.

[0156] The instructions explicitly require the model to output the extracted entities in JSON format, with each entity containing information such as name, type, and original context.

[0157] S23, Extracting Execution Relationships

[0158] After entity extraction is complete, the relation extraction agent analyzes entity pairs within the same sentence or context. First, based on the relation types defined in the conceptual model, the system filters out a set P of candidate entity pairs that meet the relation constraints. This set can be represented as:

[0159] ;

[0160] in, and Representing two entities, E type1 E type2 represents the entity type in the conceptual model, r represents the specific relationship type between entity types defined in the conceptual model, and O(R) represents the set of relationships defined in the schema.

[0161] In this way, relationship exploration is performed only among predefined combinations of entity types that may have relationships, avoiding the high computational cost of brute-force methods.

[0162] For each candidate pair, the system builds instructions that require the LLM to determine whether a relationship exists between them and the type of relationship. For example, for the entity pair ("Kingfisher's beak", "High-speed train front design"), based on the context "inspired", the model should be able to correctly extract the relation triples: (<Kingfisher's beak>, generateInnovation, <High-speed train front design>).

[0163] S3, Multi-dimensional Collaborative Knowledge Verification

[0164] like Figure 4 As shown, to ensure knowledge quality, this invention designs a multi-dimensional collaborative verification mechanism throughout the entire extraction process. Each extracted knowledge triple must pass the verification in the following three dimensions:

[0165] S31, Context Consistency Verification

[0166] This verification aims to ensure that the extracted knowledge is faithful to the original text. It is implemented by the system invoking a "contextual verification agent" based on a large language model. This agent receives a prompt containing the following:

[0167] (1) Role: "You are a meticulous text analyst."

[0168] (2) Task: "Please determine whether the following 'knowledge to be verified' can be fully supported by the 'original context'. If the knowledge is a direct restatement of the original text or a strong logical implication, it is highly consistent; if it is partially supported or weakly related, it is moderately consistent; if it is irrelevant or contradictory, it is inconsistent."

[0169] (3) Input: Knowledge to be verified (such as entity “hummingbird wings”) and original context text block.

[0170] (4) Output format: Returns a JSON object containing a "consistency score" (0.0 to 1.0) and a "validation reason" (natural language interpretation).

[0171] This verification can be represented as:

[0172] ;

[0173] in, This indicates the context consistency verification result of the entity. This indicates that the entity extracts the context text from its source, LLM. context This represents a context-verification agent based on a large language model.

[0174] S32. Conceptual Model Consistency Verification

[0175] This verification aims to ensure that all extracted knowledge strictly adheres to the domain rules defined in step S1. It is implemented by the system invoking a "Schema Verification Agent," which is also based on the large language model and receives a prompt containing the following:

[0176] (1) Role: "You are a knowledge graph architect responsible for maintaining the integrity of the schema."

[0177] (2) Task: “Based on the following ‘conceptual model definition’, determine whether the type and structure of ‘knowledge to be verified’ are fully compliant.”

[0178] (3) Input: Complete conceptual model definition (eight entity types and relational constraints) and knowledge to be verified (such as relational triples "<entity A: type X> ab <entity B: type Y>").

[0179] (4) Output format: Returns a JSON object containing a "compliance score" (0.0 or 1.0) and a "validation reason" (e.g., "The head entity type of relation ab must be Z, but it is actually X, which is non-compliant").

[0180] This verification can be represented as:

[0181] .

[0182] S33, Verification of Factual Consistency

[0183] This verification aims to leverage extensive knowledge of the external world to examine potential "illusions" arising from LLM. It is implemented by having the system invoke a "fact-verifying agent" equipped with web search tools. This agent performs the following sub-tasks:

[0184] (1) Generate search queries: Based on the knowledge to be verified (such as "hummingbird wings vibrate 80 times per second"), generate multiple search queries (such as "hummingbird wing beat frequency").

[0185] (2) Perform web search: Call a commercial search engine via API to execute a query and obtain the returned web page summary and URL.

[0186] (3) Comprehensive judgment: The agent receives a prompt containing the following:

[0187] 1) Role: "You are a rigorous scientific fact checker."

[0188] 2) Task: "Please analyze the following 'online search results' to determine the authenticity of the 'knowledge to be verified'."

[0189] 3) Input: Knowledge to be verified and all online search results.

[0190] 4) Output format: Returns a JSON object containing a "fact score" (0.0 to 1.0), a "verification reason", and a "list of source URLs" for all references.

[0191] This verification can be represented as:

[0192] .

[0193] in, e represents the result of entity fact consistency verification. i The entity represents the entity, and Tool represents the external network search tool invoked by the agent. LLM fact This represents a fact-verifying agent based on a large language model.

[0194] S4. Collaborative decision-making based on the Delphi method consultation mechanism

[0195] After the multi-dimensional verification in step S3, the system obtains a verification score (v) for a knowledge item in three dimensions: context, conceptual model, and external facts. context v schema v fact And the corresponding natural language reasoning. Traditional decision-making methods, such as simple weighted fusion, while intuitive, cannot adapt to the complexity of different knowledge items and the potential conflicts between verification dimensions due to their fixed weights. The decision-making process lacks robustness and interpretability. For example, when a knowledge item scores highly on context and schema but scores poorly on fact verification, a simple weighted average may yield an ambiguous intermediate score, making it impossible to make a clear judgment.

[0196] To overcome this limitation, this invention introduces a more advanced multi-agent collaborative decision-making mechanism based on the Delphi method. Instead of directly fusing verification scores, this mechanism constructs a committee composed of multiple decision agents (DAs) with different decision-making strategies. These DAs first independently make a comprehensive judgment on the verification results, and then, through the Delphi method consultation mechanism, they iteratively discuss and revise their respective judgments in multiple rounds until a consensus is reached. This decision-making architecture, which builds upon existing decisions, greatly improves the accuracy, robustness, and transparency of the final decision.

[0197] like Figure 5 As shown, the specific implementation steps of this collaborative decision-making mechanism are as follows:

[0198] S41. Establish a parallel decision-making intelligent agent committee.

[0199] This embodiment sets up three parallel decision agents (DA1, DA2, DA3). Each agent receives all the verification results from step S3 (i.e., three verification scores and corresponding verification reasons), and independently gives a preliminary decision score d based on its unique built-in decision strategy. j (j∈{1,2,3}) and decision reasoning R j .

[0200] S411, Conservative Agent

[0201] Decision-making strategy: The agent is based on a large language model and guided by a decision prompt that emphasizes "high standards and zero tolerance." This prompt knowledge model only gives a high score when all validation dimensions provide extremely high confidence and the reasoning is irrefutable, aiming to maximize the accuracy of decision-making.

[0202] Decision function: d1, R1 = LLM(..., Prompt_conservative);

[0203] Decision Reason R1: Generated by LLM based on conservative prompts, it clearly indicates the rigor and high standards of the decision.

[0204] S412, Balanced Agent

[0205] Decision-making strategy: This agent is also based on a large language model, but its decision prompts emphasize "comprehensive consideration and balanced evaluation." These prompts guide the model to find a balance between precision and recall, and comprehensively evaluate the scores and justifications of each validation dimension.

[0206] Decision function: d2, R2 = LLM(..., Prompt_balanced);

[0207] Decision Reason R2: Generated by LLM based on balanced prompts, it reflects the process of weighing opinions on different validation dimensions.

[0208] S413, Intelligent Agent (Decision-Making Intelligent Agent)

[0209] Decision-making strategy: This agent shares the same underlying model as the previous two, but its prompts are the most complex, emphasizing "deep insight and logical reasoning." These prompts encourage the model to deeply analyze and verify the potential conflicts or supporting relationships between the reasons, making the most intelligent judgment.

[0210] Decision function: Its decision score d3 is generated by an LLM based on a comprehensive judgment of all input information.

[0211] ;

[0212] Decision Reason R3: A detailed and interpretable decision analysis report is generated directly from the LLM, explaining the complete thought process behind the judgment.

[0213] By setting up these three decision-making agents with different "personalities," the system can examine the knowledge to be verified from three different perspectives: conservative, balanced, and intelligent, providing rich and differentiated initial opinions for subsequent Delphi consultations.

[0214] S42, Implementation of the Delphi Law Consultation

[0215] After the three decision agents give their initial judgments, the system initiates the Delphi consultation process to discuss and revise their decision results (d1, d2, d3) in multiple rounds.

[0216] S421, Initial Consultation

[0217] Set the maximum number of consultation rounds N max (In this embodiment, it is set to 3) and the threshold θ for opinion convergence. convergence (In this embodiment, it is set to 0.05). Initialize the consultation round k=1, and use the initial decision scores of each decision agent as the opinions of the first round: d j 1 =d j .

[0218] S422, Multiple rounds of anonymous feedback and revisions

[0219] At the start of the k-th round of consultation (k>1), the system first distributes an anonymous feedback report F to each decision agent. (k-1) The report includes the decision scores, rationale, and statistical information (such as mean, variance, etc.) of all decision agents in the previous round (round k-1), but does not specify the specific source of each opinion.

[0220] ;

[0221] in, This represents the anonymous feedback report generated in the (k-1)th round; Aggregate represents the information aggregation function, used to organize and summarize the outputs of each agent; , , These represent the initial decision scores given by the three decision-making agents—conservative, balanced, and intelligent—in the (k-1)th round, respectively. , , These represent the specific reasons or arguments given by the three decision-making agents in the (k-1)th round.

[0222] After receiving the feedback report, each decision-making agent will re-examine its judgment and adjust it based on the group's opinion, generating a new round of decision scores d. j (k) Its update function can be expressed as:

[0223] ;

[0224] in, and Let represent the decision score and decision reasoning output by the j-th decision agent in the current k-th round after adjustment, respectively; This represents the update policy function unique to the j-th agent, used to perform cognitive iteration by combining its own historical judgments with group feedback; and This indicates the decision state of the agent in the previous round (round k-1); This is the anonymous feedback report received from the previous round.

[0225] For example, if the conservative agent DA1 gives a very low score in the first round, but sees in the feedback report that the other two agents gave high scores and sufficient reasons, it might moderately increase its score in the second round, or add in its reasons that although it adheres to a conservative judgment, it has noticed that the group opinion tends to accept it and suggests further review. This mechanism simulates the process by which human expert groups inspire each other and gradually converge in discussions.

[0226] S423. Computational Consistency and Convergence Judgment

[0227] After each round of consultation, the system calculates the consistency index of the opinions of each decision-making agent in the current round. This embodiment uses the coefficient of variation (CV) as a consistency metric because it can better handle comparisons of dispersion under different means.

[0228] The consistency index The calculation formula is as follows:

[0229] ;

[0230] To calculate the aforementioned coefficient of variation, it is necessary to calculate the decision mean for each round. and standard deviation The calculation formulas are as follows:

[0231] ;

[0232] ;

[0233] in, The coefficient of variation represents the results of the k-th round of consultations, serving as a core indicator for measuring the consistency of expert opinions in that round. The standard deviation of the decision scores of all decision agents in the k-th round is used to reflect the dispersion of the decision scores. This represents the arithmetic mean of the decision scores of all decision agents in the k-th round; The value represents the decision score given by the j-th decision agent in the k-th round; M represents the total number of decision agents participating in the consultation, which is set to 3 in this embodiment; j represents the index number of the decision agent, ranging from 1 to M; k represents the current consultation iteration round.

[0234] The system according to The value of the convergence threshold is used to perform convergence judgment logic, which involves two key preset hyperparameters: convergence threshold and convergence threshold. and maximum number of consultation rounds .in, This represents the maximum allowable dispersion for the system to determine that expert opinions have reached a consensus; in this embodiment, it is set to 0.05. This represents the maximum number of consultation iterations set to prevent infinite loops; in this embodiment, it is set to 3. The specific judgment and execution logic is as follows:

[0235] if Once it is determined that the opinions of the decision-making agents have reached a consensus, the consultation has successfully converged, the process ends, and the process proceeds to S43 for final decision aggregation.

[0236] if and The consensus is not yet reached and further discussion is needed. Let k = k + 1, and return to S42B for the next round of consultation.

[0237] if and The maximum number of consultation rounds has been reached, but significant differences of opinion still exist. This usually indicates that the knowledge item itself is highly controversial or ambiguous. In this case, the consultation should be forcibly ended, and the discussion should proceed to S44 for dispute resolution.

[0238] S43, Consensus Decision Aggregation

[0239] When the consultation successfully converges, the system performs a final round (round K, K, N). max The decision scores are aggregated to calculate the final effectiveness score v. final This embodiment uses a weighted average method, but the weights will be dynamically adjusted based on the consistency of the round, giving higher weight to opinions that are closer to the group mean.

[0240] ;

[0241] in, This represents the final determined score for the validity of the knowledge. This represents the total number of decision-making agents; Let $\frac{j}{j}$ represent the decision score given by the j-th decision agent in the K-th round. This represents the aggregate weight of the J-th decision agent in the K-th round. The weight... The value of is related to the score of the agent. It is inversely proportional to the absolute difference between the mean of the current group, that is... The closer to the mean, the higher the corresponding weight. The higher the value, the less the impact of outlier opinions on the final result.

[0242] S44, Disagreement Handling and Manual Review Marking

[0243] If the consultation fails to converge after reaching the maximum number of rounds, the system will no longer forcibly calculate a composite score, but will instead mark the knowledge item as "For Manual Review." Simultaneously, the system will generate a detailed "Disagreement Report," which includes:

[0244] (1) Original verification results (three verification dimensions).

[0245] (2) The decision scores and reasons for each decision by the three decision agents in each round.

[0246] (3) Changes in consistency indicators in each round.

[0247] This report provides comprehensive information on the decision-making process for domain experts to review, enabling them to quickly identify points of disagreement and make a final decision. This "disagreement to human" mechanism ensures that the system does not make hasty or erroneous judgments when faced with complex and ambiguous knowledge, demonstrating the intelligence of human-machine collaboration.

[0248] S45. Threshold Judgment and Knowledge Acceptance

[0249] The final validity score v is obtained through S43 consensus decision aggregation. final The system compares the knowledge entries with a preset acceptance threshold T. final (In this embodiment, it is 0.85) for comparison. If v final ≥T final If the answer is yes, the knowledge is accepted and proceeds to the subsequent S5 process; otherwise, it is rejected.

[0250] Through the aforementioned multi-agent collaborative decision-making mechanism based on the Delphi method, this invention upgrades the decision-making process from a simple "calculation problem" to a dynamic and interpretable "debate process," significantly improving the intelligence level of knowledge graph construction and the reliability of the final knowledge. In practice, this invention has observed that most knowledge items converge after 2-3 rounds of consultation, while a few complex or ambiguous knowledge items may require 4-5 rounds.

[0251] Advantages of the Delphi Consultation Mechanism

[0252] Compared to traditional hard-weighted decision-making, the Delphi method consultation mechanism has the following significant advantages:

[0253] (1) Dynamic adjustment: Each verification agent can dynamically adjust its judgment based on the opinions of other agents, thus avoiding the rigidity of fixed weights.

[0254] (2) Information interaction: Through multiple rounds of feedback, the agents achieve full information interaction, which can correct and supplement each other, thus improving the comprehensiveness of the verification.

[0255] (3) Consistency guarantee: By setting a convergence threshold, the system can ensure that the final decision is made on the basis of a basic consensus in each verification dimension, which improves the reliability of the decision.

[0256] (4) Explainability: Each round of consultation recorded the verification reasons of each agent, providing a detailed explanation for the final decision, which facilitates manual review and system debugging.

[0257] S5, Knowledge Alignment, Integration and Storage

[0258] S51, Entity Alignment and Blending

[0259] After multi-dimensional verification and Delphi method consultation and decision-making, the system obtains a batch of high-quality entities and relationships. However, due to the diversity of source texts and differences in expression, there may be duplicate or equivalent entries in this knowledge. For example, entities such as "hummingbird's wing," "hummingbird wing," and "wing of hummingbird" may be extracted from different documents, all of which actually refer to the same concept. The entity alignment module is responsible for identifying and merging these equivalent entities.

[0260] like Figure 6 As shown, the detailed implementation of entity alignment is as follows:

[0261] S511, Entity Vectorization: For each verified entity, the system collects the following information: entity name, entity type, and all relevant entity attributes and their values, i.e., the attribute values ​​of the entity that conform to the schema in all extraction results. Then, this string is input into a powerful language model (this embodiment uses OpenAI's text-embedding-3-large model, which generates 3072-dimensional embedding vectors) to generate a semantic vector representation v for the entity. entity This vector captures the entity's location in a multidimensional semantic space, containing comprehensive semantic information about its name, type, and attributes.

[0262] S512. Similarity Calculation and Clustering: After collecting the semantic vectors of all entities, the system uses an iterative clustering algorithm based on a similarity threshold to find semantically similar entities. The cosine similarity calculation formula is as follows:

[0263] ;

[0264] in, and Let represent the semantic vectors of the two entities, · denotes the vector dot product, and ||·|| denotes the L2 norm of the vector.

[0265] The clustering algorithm process is as follows: The time complexity of this algorithm is O(n²), where n is the number of entities. For large-scale knowledge graphs (such as those containing hundreds of thousands of entities), Approximate Nearest Neighbor (ANN) algorithms (such as the FAISS library) can be used to accelerate similarity calculation, reducing the complexity to O(nlogn). When sim(v i v j )≥T align When the value is typically set to 0.90-0.95, the two entities are considered equivalent and are merged.

[0266] S513. Canonical Naming: For each cluster (representing a unique concept), the system needs to select a canonical name for it. This embodiment adopts the following strategy: (1) Count the frequency of all entity names in the cluster (based on their number of occurrences in the original text). (2) Select the name with the highest frequency as the canonical name. (3) If there are multiple names with the same frequency, select the shortest one (assuming that a more concise name is more canonical). (4) Store other names as aliases for the concept.

[0267] S514, Entity ID Assignment: Assign a globally unique entity ID (such as a UUID or an auto-incrementing integer) to each unique concept. All subsequent knowledge triples involving this concept will use this ID, instead of the original text name, thereby achieving true deduplication and unification.

[0268] S52, Relationship Alignment and Blending

[0269] Similar to entity alignment, there can also be redundancy between relation triples. For example, (<hummingbird>, haveFunction, <hover>) and (<hummingbird>, haveFunction, <hover capability>) should also be considered duplicates if "hover" and "hover capability" are merged into the same entity during the entity alignment phase.

[0270] The system normalizes all relation triples by replacing the entity names in the triples with their corresponding normalized entity IDs. Then, the triples are deduplicated. For identical triples (with the same head entity ID, relation type, and tail entity ID), the system retains one and merges its source information (such as the list of source documents and the number of extractions) as additional evidence of the relation's reliability.

[0271] S53. Knowledge is stored in the Neo4j graph database.

[0272] This embodiment selects Neo4j as the graph database. Neo4j is a high-performance, native graph database that supports ACID transactions, the Cypher query language, and a rich library of graph algorithms, making it very suitable for storing and querying knowledge graphs.

[0273] S531. Database connection and configuration: The system uses the Neo4j Python driver (neo4j library, version 5.9.0) to connect to the Neo4j server.

[0274] S532, Node Creation: For each aligned, unique entity, the system creates a node in Neo4j. The node's label corresponds to its entity type in the schema (e.g., :BioEntity, :Function). The node's properties store detailed information about the entity.

[0275] The Cypher query for creating a node is as follows:

[0276] CREATE (e:BioEntity {id: 'uuid-12345', name: 'hummingbird', aliases: ['hummingbird', 'hummingbird family'], description: 'A small bird known for its unique hovering ability and rapid wingbeats.', sources:['https: / / en.wikipedia.org / wiki / Hummingbird', 'https: / / pubmed.ncbi.nlm.nih.gov / ...']})

[0277] The system improves write efficiency through batch operations. For example, the UNWIND statement can be used to create multiple nodes at once.

[0278] S533, Relation Creation: For each validated relation triple (head_id, relation_type, tail_id), the system creates a directed edge between the corresponding two nodes in Neo4j. The edge type (Type) is the relation type (relation_type) (e.g., HAVE_FUNCTION, GENERATE_INNOVATION). The edge's attributes can store metadata about the relation, such as source text, confidence score, number of extractions, etc. The Cypher query for creating the relation is as follows:

[0279] MATCH (h {id: 'uuid-12345'}), (t {id: 'uuid-67890'})

[0280] CREATE (h)-[r:HAVE_FUNCTION{confidence: 0.92, sources: ['Document A', 'Document B'], extraction_count: 5}]->(t)

[0281] S534, Indexes and Constraints: To speed up queries, the system creates indexes on frequently used attributes. For example, indexes are created on the id and name attributes of all nodes:

[0282] CREATE INDEX entity_id_index FOR (e:Entity) ON (e.id);

[0283] CREATE INDEX entity_name_index FOR (e:Entity) ON (e.name);

[0284] In addition, to ensure data integrity, the system creates uniqueness constraints for entity IDs:

[0285] CREATE CONSTRAINT entity_id_unique FOR (e:Entity) REQUIRE e.id ISUNIQUE;

[0286] S535, Transaction Management: All database write operations are executed within a transaction to ensure ACID properties. If a batch of writes fails, the transaction is rolled back to prevent data inconsistency.

[0287] S54. Visualization and Query of Knowledge Graphs

[0288] Once built, users can visualize and query the knowledge graph through the Neo4j Browser or a custom web interface. For example, to query all biomimetic innovations related to "hummingbird":

[0289] MATCH (b:BioEntity {name:'Hummingbird'})-[r*1..3]-(i:Innovation);

[0290] RETURN b, r, i.

[0291] This query returns all Innovation nodes reached from the hummingbird node through 1 to 3 hops, thus revealing the various technological innovations inspired by hummingbirds.

[0292] To comprehensively evaluate the beneficial effects of this invention, this embodiment conducted detailed experiments on a gold standard test set containing 1700 samples. This test set was manually annotated by three bionics experts, containing 850 entity-annotated samples and 850 relation-annotated samples. Evaluation metrics included standard precision (P), recall (R), and F1 score.

[0293] Experiment 1: Overall Performance Evaluation

[0294] The MAG (Multi-Agent Graph Construction) method proposed in this invention achieves excellent performance, as shown in the table below:

[0295]

[0296] The formulas for calculating precision (P), recall (R), and F1 score are as follows:

[0297] ;

[0298] Wherein, TP (True Positive) represents the number of correctly extracted knowledge points, FP (False Positive) represents the number of incorrectly extracted knowledge points, and FN (False Negative) represents the number of missed knowledge points.

[0299] The formula for calculating entity redundancy rate (ERR) is:

[0300] ;

[0301] As shown in the table above, the F1 score for entity extraction reaches 0.662, and the F1 score for relation extraction is even higher at 0.907, which is a very advanced level in the field of automated knowledge graph construction. The accuracy of relation extraction is close to 94%, indicating that the multi-dimensional verification mechanism of this invention effectively suppresses the illusion of LLM and ensures high-quality extracted knowledge. Schema consistency reaches 100%, proving the effectiveness of the conceptual model constraint in this invention, and all extracted knowledge strictly conforms to predefined domain rules. The entity redundancy rate (ERR) is as low as 5.6%, proving the effectiveness of the entity alignment module, and the entities in the knowledge graph are highly deduplicated and normalized.

[0302] Experiment 2: Comparison with baseline methods

[0303] To highlight the advantages of this invention, it is compared with a variety of baseline methods, including traditional knowledge extraction methods, general LLM direct extraction, and methods that use only fine-tuned LLM without verification mechanisms.

[0304]

[0305] As shown in the table above, compared to the traditional method DeepKE, this invention improves the F1 score for entities by 45.6% and for relations by 52.0%, demonstrating the significant advantages of LLM-based methods in knowledge extraction tasks. Compared to the powerful general-purpose LLM GPT-4o (zero-shot), this invention improves the F1 score for relation extraction by nearly 15 percentage points (from 0.759 to 0.907), demonstrating the significant value of domain fine-tuning and verification mechanisms. Compared to methods that only fine-tune without verification, this invention improves the F1 score for relations by 14.1% (from 0.795 to 0.907), highlighting that multi-dimensional collaborative verification and Delphi method consultation are key to ensuring knowledge quality.

[0306] Experiment 3: Verification Module Ablation Experiment

[0307] To evaluate the contribution of each verification dimension separately, this invention conducted an ablation study, in which one or two of the context verification, schema verification, and fact verification were removed, and the performance changes were observed.

[0308]

[0309] As shown in the table above, removing any one of the verification modules leads to a significant performance decrease. Removing context verification has the greatest impact on recall (from 0.878 to 0.753), while removing fact verification and schema verification has the greatest impact on precision. Removing two verification modules results in an even more pronounced performance decrease, with the F1 score dropping by up to 12.8 percentage points (from 0.907 to 0.779). This strongly demonstrates that in the multi-dimensional verification mechanism designed in this invention, each dimension is indispensable, and they collectively ensure the quality of knowledge from different perspectives.

[0310] Experiment 4: Effectiveness Assessment of the Delphi Method Consultation Mechanism

[0311] To evaluate the advantages of the Delphi consultation mechanism over traditional hard-weighted decision-making, this invention conducted a comparative experiment.

[0312]

[0313] As shown in the table above, the Delphi method's consultation decision-making improved the F1 score for relation extraction by 1.3 percentage points (from 0.894 to 0.907) compared to the hard-weighted decision-making method, demonstrating the effectiveness of the consultation mechanism. The average number of consultation rounds is 2.3, indicating that most knowledge items can converge after 2-3 rounds of consultation, and the computational cost of the system is acceptable.

[0314] Further analysis revealed that the Delphi method consultation mechanism has a more obvious advantage when dealing with complex or ambiguous knowledge items. It can gradually eliminate differences and reach more accurate verification conclusions through multiple rounds of information exchange.

[0315] Experiment 5: The Influence of Entity Alignment Threshold

[0316] To evaluate the entity alignment threshold T align To assess the impact on the final knowledge graph quality, this invention conducted a series of experiments, testing different T values. align The entity redundancy rate (ERR) and recall rate (R) are calculated at these values.

[0317]

[0318] The results show that T alignAt a threshold of 0.95, the system achieves the optimal balance between entity redundancy and recall. A threshold that is too low can lead to over-merging, incorrectly classifying different entities as equivalent, thus reducing recall; a threshold that is too high can lead to under-merging, resulting in a large number of equivalent entities not being identified, and increasing redundancy.

[0319] The detailed technical descriptions and quantified experimental results above fully demonstrate that the proposed method for automatic construction of biomimetic domain knowledge graphs based on multi-agent systems is not only technically complete and feasible, but also has significant and quantifiable advantages in practical effects compared to existing technologies. It successfully solves the various challenges raised in the background technology and provides a powerful knowledge infrastructure for biomimetic research and application.

[0320] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0321] This embodiment describes in detail how to use LoRA technology to fine-tune a pre-trained large language model to enable it to possess professional knowledge in the field of bionics.

[0322] The mathematical principles of LoRA technology:

[0323] Traditional full fine-tuning requires updating all parameters of the model. For a large language model with billions of parameters, this not only requires enormous computational resources and GPU memory but also easily leads to overfitting. LoRA technology, by introducing low-rank decomposition, significantly reduces the number of trainable parameters.

[0324] For a weight matrix W0∈R in the pre-trained model (d×k) LoRA does not directly update W0, but keeps it frozen and adds a low-rank update ΔW next to it:

[0325] ΔW = B·A;

[0326] Where A∈R (r×k) , B∈R (d×r) ,r< <min(d,k)。

[0327] During forward propagation, the calculation of output h becomes:

[0328] ;

[0329] The number of trainable parameters is reduced from d×k to r*(d+k). For example, for a Transformer layer with d=k=4096, if full parameter fine-tuning is used, 4096×4096=16,777,216 parameters need to be trained; while using LoRA with r=16, only 16×(4096+4096)=131,072 parameters need to be trained, a reduction of approximately 99.2%.

[0330] To control the influence intensity of the LoRA layer, a scaling factor α is introduced:

[0331] ;

[0332] During initialization, matrix A is usually initialized with Gaussian randomization, while matrix B is initialized as a zero matrix. This way, ΔW=0 at the start of training, and the initial behavior of the model is completely consistent with the pre-trained model.

[0333] Detailed construction of the fine-tuning dataset

[0334] The biomimetic instruction fine-tuning dataset constructed in this embodiment mainly contains entity extraction samples. These entity extraction samples refer to a given biomimetic-related text where the model is required to extract all entities and their types. Each sample contains three fields: instruction, input, and output. To save storage space and improve training efficiency, long texts have been simplified. An example is shown below:

[0335] (1) "instruction": "You are a professional intelligent agent in the field of biology, focusing on the construction of knowledge graphs. Please extract entities that meet the definitions from the input text according to the following predefined schema (including entity types such as Challenge, Innovation, BioStrategy and their attribute definitions). Please ensure that your answer is rigorous and output the results in the specified JSON format. If there are no matching entities, leave it blank."

[0336] (2) "input": "The mussel mantle attracting larval hosts is a unique biological strategy that demonstrates the complexity of mutualistic relationships in nature. For example, the smooth-shelled pygmy mussel achieves stability and protection in water through its shell structure. The functional design of this structure provides an important reference for biomimetic, especially in the development of materials and technologies that can withstand high flow rates or high pressure environments."

[0337] (3) "output":"{"entities":[{"name":"Mussel mantle attracts larval hosts","type":"BioStrategy","properties":{"name":"Mussel mantle attracts larval hosts","description":"A strategy to attract hosts by mimicking predator morphology to achieve offspring propagation"}},{"name":"Smooth miniature mussel","type":"BioEntity","properties":{"name":"Smooth miniature mussel","description":"A freshwater mussel with a special shell structure"}},{"name":"Shell","type":"BioComponent","properties":{"name":"Shell"}}]}"

[0338] During training, the system monitors the following metrics in real time:

[0339] (1) Training Loss: The cross-entropy loss function is used to measure the difference between the model's prediction and the true label. The training loss should decrease monotonically with the number of training steps.

[0340] (2) Validation Loss: The loss calculated on the validation set and used to monitor overfitting. If the validation loss starts to rise while the training loss continues to fall, it indicates that the model is starting to overfit.

[0341] (3) Learning Rate: A linear warm-up + linear decay strategy is adopted. In the first 10% of training steps, the learning rate increases linearly from 0 to 2×10^(-4); thereafter it decreases linearly to 0.

[0342] (4) Gradient Norm: Monitors the L2 norm of the gradient to detect gradient explosion or gradient vanishing problems. In this embodiment, the gradient norm is clipped to the maximum value of 1.0.

[0343] The training curves show that the training loss of the model tends to stabilize after about 5,000 steps, and the validation loss reaches its lowest point at about 8,000 steps, after which it rises slightly. Therefore, the checkpoint at the 8,000th step was selected as the final model.

[0344] To quantify the effect of fine-tuning, this embodiment compares the performance of the model before and after fine-tuning on the test set:

[0345]

[0346] The results show that fine-tuning brings significant performance improvements to the entity extraction task (F1 score improved from 0.372 to 0.681, an improvement of 83.1%). This is because entity extraction is highly dependent on domain knowledge, and the un-fine-tuned model struggles to accurately recognize biomimetic terminology. Fine-tuning also brings a 5.5% improvement to the relation extraction task, demonstrating the value of domain adaptation.

[0347] The detailed embodiments described above fully demonstrate that the technical solution of the present invention is complete, feasible, and has significant advantages, providing an efficient, accurate, and scalable automated solution for the field of bionics and even broader knowledge graph construction tasks.

[0348] Example 2

[0349] In a second aspect of the invention, a biomimetic domain knowledge graph automatic construction system based on multi-agent systems is disclosed, the system comprising:

[0350] The concept model building module is used to build concept models in the field of bionics.

[0351] The knowledge extraction module is used to construct a multi-agent system based on the bionic domain concept model, and use the multi-agent system to automatically extract the semantic relationships between bionic entities from unstructured text, and output the entity set and relation triplet that conform to the model specification.

[0352] The knowledge verification module is used to perform multi-dimensional collaborative verification of the biomimetic entity and the semantic relationships between entities using the multi-agent system;

[0353] The Delphi method consultation and decision-making module is used to make collaborative decisions on the verification results based on the Delphi method consultation mechanism. After multiple rounds of anonymous feedback and opinion revision, a consensus is reached, and a final judgment is made on the validity of the knowledge.

[0354] The knowledge graph generation module is used to align and merge verified entities and relationships, store them in a graph database, and form a structured biomimetic domain knowledge graph.

[0355] Example 3

[0356] An electronic device includes: at least one processor; and a memory storing instructions that, when executed by the at least one processor, cause the at least one processor to perform the above-described method for automatically constructing a biomimetic domain knowledge graph based on multiple agents.

[0357] In this embodiment, the electronic device may include, but is not limited to: personal computer, server computer, workstation, desktop computer, laptop computer, notebook computer, mobile computing device, smartphone, tablet computer, cellular phone, personal digital assistant (PDA), handheld system, messaging device, wearable computing device, consumer electronic device, etc.

[0358] Example 4

[0359] A machine-readable storage medium storing executable instructions that, when executed, cause the machine to perform the above-described biomimetic domain knowledge graph automatic construction method based on multiple agents.

[0360] Specifically, a system or system equipped with a readable storage medium may be provided, on which software program code implementing the functions of any of the embodiments described above is stored, and the computer or processor of the system or system can read and execute the instructions stored in the readable storage medium.

[0361] In this case, the program code read from the readable medium itself can perform the functions of any of the above embodiments, and therefore the machine-readable code and the readable storage medium storing the machine-readable code constitute a part of this specification.

[0362] Examples of readable storage media include floppy disks, hard disks, magneto-optical disks, optical disks (such as CD-ROM, CD-R, CD-RW, DVD-ROM, DVD-RAM, DVD-RW, DVD-RW), magnetic tapes, non-volatile memory cards, and ROMs. Alternatively, program code can be downloaded from a server computer or the cloud via a communication network.

[0363] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0364] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A system that specifies functions in one or more boxes.

[0365] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including an instruction set implemented in a process. Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0366] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0367] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A method for automatically constructing a biomimetic domain knowledge graph based on multi-agent systems, characterized in that, The method includes the following steps: S1. Construct a conceptual model for the field of bionics; S2. Based on the bionic domain conceptual model, construct a multi-agent system, and use the multi-agent system to automatically extract the semantic relationships between bionic entities from unstructured text, and output the entity set and relation triplet that conform to the model specification. S3. Utilize the multi-agent system to perform multi-dimensional collaborative verification of the biomimetic entity and the semantic relationships between entities; S4. Based on the Delphi method consultation mechanism, collaborative decision-making is carried out on the verification results. After multiple rounds of anonymous feedback and opinion revision, a consensus is reached, and a final judgment is made on the validity of the knowledge. S5. Align and merge the verified entities and relationships, and store them in a graph database to form a structured biomimetic domain knowledge graph.

2. The method for automatic construction of biomimetic domain knowledge graphs based on multi-agent systems according to claim 1, characterized in that, Step S1 specifically includes: S11. Define the core entity type; The conceptual model in the field of bionics predefines core entity types and relation types. The core entity types include: biological entities, biological components, bionic strategies, functions, innovation principles, innovation, challenges, and application potential. The conceptual model is represented as a graph structure O=(C,R), where C is the set of entity types and R is the set of relation types. S12, Define the relation type; The relationship type is used to define various semantic relationships between entities.

3. The method for automatic construction of biomimetic domain knowledge graphs based on multi-agent systems according to claim 2, characterized in that, Step S2 specifically includes: S21. Construct a multi-agent system; the multi-agent system includes an entity extraction agent and a relation extraction agent, wherein each agent is obtained by fine-tuning a pre-trained large language model through low-rank adaptive technology to possess professional knowledge in the field of bionics. S22. Perform entity extraction; the entity extraction agent is responsible for identifying and extracting the eight entity types defined in step S11 from the input text. This extraction process can be represented as follows: ； Where E represents the extracted entity set, x is the input free text, O represents the biomimetic domain concept model, and LLM... LoRA This represents the large model after LoRA fine-tuning; S23. Perform relation extraction: The relation extraction agent first selects a set of candidate entity pairs P that meet the relation constraints based on the relation types defined in the biomimetic domain conceptual model. This process first generates a set of candidate entity pairs P based on the entity type constraints defined in the conceptual model, then performs relation judgment on the candidate entity pairs, and outputs the relation triplet to be verified. The set of candidate entity pairs P is shown in the following equation: ； in, and Representing two entities, , Let represent the entity types in the conceptual model, and r represent the specific relationship type between the entity types defined in the conceptual model. This represents the set of relationships defined in the schema.

4. The method for automatic construction of biomimetic domain knowledge graphs based on multi-agent systems according to claim 1, characterized in that, Step S3 specifically includes: S31. Perform context consistency verification on the biomimetic entities and semantic relationships extracted in step S2: Calculate the semantic similarity between the knowledge to be verified and the original unstructured text context using a semantic similarity calculation model, determine the correlation between the two, and output the context consistency verification result. For entity e i Context consistency verification is performed using the following formula: ； in, This indicates the entity context consistency verification result. The context text representing the source of entity extraction. This represents a context-verification agent based on a large language model; For relation triples (e i e j ,r), use the following formula to perform context consistency verification: ； in, and Representing two entities, Indicates a relationship; S32. Utilize a multi-agent system to verify the consistency of the conceptual model between the biomimetic entities and their semantic relationships extracted in step S2: This is done by checking whether the extracted entity types and relationship types conform to the predefined constraints in the biomimetic domain conceptual model. This process can be represented as: ； in, This indicates the entity schema consistency verification result. For a schema verification agent based on a large language model, v schema This represents the entity schema consistency verification result, and O represents the conceptual model in the biomimetic domain. S33. Use a multi-agent system to verify the factual consistency of the biomimetic entities and semantic relationships extracted in step S2: By calling the external network search tool Tool, the extracted knowledge is fact-checked and confirmed. For entities The following formula is used to verify the consistency of facts: ； in, This indicates the result of the entity fact consistency verification. This refers to the external network search tools invoked by the intelligent agent. A fact-verifying agent based on a large language model, equipped with web search tools; For relation triples (e i e j The following formula is used to verify the consistency of facts: ； in, LLM represents the result of the entity relationship factual consistency verification. fact A fact-verifying agent based on a large language model, equipped with web search tools.

5. The method for automatic construction of biomimetic domain knowledge graphs based on multi-agent systems according to claim 1, characterized in that, Step S4 specifically includes: S41. Initialize the expert agent group: Set up multiple parallel decision agents of conservative, balanced and intelligent types. Each decision agent receives all the verification results from step S3 and makes an independent comprehensive judgment based on its own preset decision strategy, outputting a preliminary decision score and decision reason; each expert agent gives an independent preliminary judgment based on different verification dimensions or domain knowledge background. S42. Multi-round anonymous consultation and iteration: The Delphi method is applied to the preliminary decision results of multiple decision agents for consultation. Through multiple rounds of anonymous feedback and correction judgment, a consensus is gradually reached. In each round of consultation, each decision agent receives the anonymous decision results and reasons of other agents, and adjusts its own judgment according to the group opinion, dynamically adjusting its own confidence and judgment conclusion, and completing multiple rounds of correction and iteration. S43. Consistency Judgment and Iteration Termination: After each round of consultation, the consistency index of expert opinions is calculated. If the consistency reaches the preset threshold, or the number of iterations reaches the maximum preset value, the iteration is terminated. S44. Final Decision Formation: A final decision is formed based on the consultation results. If the opinions of the decision-making agents converge, the final validity score is formed by aggregating the expert judgment results of the final round through a voting mechanism or weighted average method. If the opinions of the decision-making agents do not converge, the knowledge is marked as pending manual review. At the same time, a detailed report containing the decision-making process of each round and the core points of disagreement is automatically generated to provide a basis for manual review. The final determination of the validity of the knowledge is completed in sequence.

6. The method for automatic construction of biomimetic domain knowledge graphs based on multi-agent systems according to claim 5, characterized in that, In step S42, the expert agent's judgment adjustment process adopts a feedback-based dynamic adjustment mechanism, whereby each expert agent updates its confidence level based on the opinions of other experts.

7. The method for automatic construction of biomimetic domain knowledge graphs based on multi-agent systems according to claim 5, characterized in that, In step S43, the consistency index is measured using the coefficient of variation (CV).

8. The method for automatic construction of biomimetic domain knowledge graph based on multi-agent technology according to claim 1, characterized in that, Step S5 specifically includes: S51. Knowledge Alignment and Fusion: Using a semantic vector-based approach, all extracted entities are encoded into high-dimensional semantic vectors through a language model. By calculating the cosine similarity between vectors, entities pointing to the same concept are automatically identified and merged to eliminate redundancy. The formula for calculating the cosine similarity is: ； in, This represents the similarity between two vectors. and These represent the semantic vectors of the two entities, Represents the vector dot product. The L2 norm of a vector is represented; the extracted entities are encoded into semantic vectors, and the cosine similarity between vectors is calculated to identify and merge duplicate or equivalent knowledge entries to eliminate knowledge redundancy. S52. Knowledge Storage: This involves storing aligned and integrated knowledge... knowledge The data is loaded into a graph database for persistent storage. Entities are stored as nodes in the graph database, and the semantic relationships between entities are stored as edges between nodes, ultimately forming a structured biomimetic domain knowledge graph.

9. The method for automatically constructing a biomimetic domain knowledge graph based on multi-agent systems according to claim 1, characterized in that, The graph database is Neo4j, which is used to store and manage entities, relationships, and their attributes.

10. A biomimetic domain knowledge graph automatic construction system based on multi-agent intelligence, characterized in that, The system includes: The concept model building module is used to build concept models (Schema) in the field of bionics. The knowledge extraction module is used to construct a multi-agent system based on the bionic domain concept model, and use the multi-agent system to automatically extract the semantic relationships between bionic entities from unstructured text, and output the entity set and relation triplet that conform to the model specification. The knowledge verification module is used to perform multi-dimensional collaborative verification of the biomimetic entity and the semantic relationships between entities using the multi-agent system; The Delphi method consultation and decision-making module is used to make collaborative decisions on the verification results based on the Delphi method consultation mechanism. After multiple rounds of anonymous feedback and opinion revision, a consensus is reached, and a final judgment is made on the validity of the knowledge. The knowledge graph generation module is used to align and merge verified entities and relationships, store them in a graph database, and form a structured biomimetic domain knowledge graph.

Images