DeepSeek-based aquatic disease prevention and treatment knowledge graph dynamic completion method, electronic equipment and storage medium

Abstract

The invention provides a DeepSeek-based dynamic completion method for an aquatic product disease prevention and control knowledge graph, electronic equipment and a storage medium, and belongs to the field of aquatic product disease prevention and control. According to the method, firstly, an entity classification labeling strategy based on a harmful object is designed aiming at the phenomenon of'different measures against the same disease ', a DeepSeek model oriented to vertical field prompt template optimization is introduced, and a global-local dual-order collaborative enhancement retrieval mechanism is combined, so that the problem of context dependence failure of a long text is effectively relieved, and the precision and consistency of semantic retrieval are improved; and secondly, in order to enhance the understanding of the deep semantic relationship among the triples with the phenomenon of'different measures from the same disease ', a micrologic rule enhancement module is adopted, expert knowledge is fused into neural representation learning in a symbol logic form, the completion accuracy is improved, and the interpretability of the model is enhanced. According to the method, deep completion and reasoning optimization of the knowledge graph are realized, and targeted prevention and treatment schemes can be provided for different aquatic animals.

Description

Technical Field

[0001] This invention relates to the field of aquatic disease prevention and control technology, and in particular to a method for dynamically completing a knowledge graph of aquatic disease prevention and control based on DeepSeek, an electronic device, and a storage medium. Background Technology

[0002] Disease prevention and control is a crucial link in ensuring the healthy development of aquaculture. Through scientific diagnosis, precise medication, and timely early warning, it effectively controls disease transmission, reduces industry losses, and improves the quality and safety of aquatic products. Knowledge graphs, as an effective tool for organizing and utilizing massive amounts of knowledge, provide a structured and systematic cognitive foundation for disease prevention and control, supporting intelligent disease diagnosis, precise medication, and risk early warning, thereby improving control efficiency. In the construction and optimization of knowledge graphs, knowledge graph completion plays a key role. Its goal is to predict and complete missing triples to enhance the completeness, consistency, and reasoning ability of the knowledge base, and it has become a core support for promoting the development of artificial intelligence applications such as information retrieval and intelligent question answering. In recent years, knowledge graph completion technology has continued to develop, and the use of large language models to complete knowledge graph completion tasks has also shown great potential.

[0003] However, despite the significant achievements of knowledge graph completion in general domains (such as search engines and recommendation systems), it still faces a series of severe challenges in vertical fields such as aquatic disease prevention and control. Among these challenges, the complexity and dynamism of the "different strategies for the same disease" phenomenon in aquatic disease data has a particularly prominent impact on the effectiveness of knowledge graph completion. For example, gill rot exhibits fundamental differences in symptoms and corresponding control strategies across different hosts (such as grass carp and shrimp) and at different times (such as different epidemic years). Spatially, the typical symptoms, epidemiological patterns, and control measures of the same disease may differ significantly depending on the target organism; failure to differentiate between them can easily lead to mismatched control recommendations. Temporally, as research progresses and pathogens evolve, the understanding and control strategies for the same disease are constantly updated; if the knowledge graph fails to reflect this in a timely manner, it will result in outdated or erroneous recommendations. Traditional knowledge representation and completion methods struggle to effectively characterize such fine-grained, temporal differences, leading to knowledge confusion and reasoning biases, which has become a major bottleneck in the construction and completion of current aquatic disease knowledge graphs.

[0004] Furthermore, existing knowledge graph completion technologies struggle to effectively handle long texts and unstructured data with dense technical terms, such as fisheries academic papers, and traditional knowledge graph completion methods have shortcomings in open-domain scenarios. Specific problems include: models based on the closed-world assumption cannot identify and link new entities; there is a lack of fine-grained semantic disambiguation mechanisms for issues such as "different solutions for the same problem"; existing open-domain knowledge graph completion methods, such as those based on fixed window masks, simple text aggregation, interactive attention, or multi-hop neighborhood fusion, generally suffer from incomplete semantic information capture, excessive computational complexity, sensitivity to text noise, and inaccurate and time-consuming extraction of relational facts when applied to long texts at the paper level, thus limiting the accuracy, reliability, and domain adaptability of knowledge extraction.

[0005] In summary, existing open-domain knowledge graph completion models are typically applied to short text scenarios and have not fully considered the impact of "different solutions for the same disease" on the completion effect of knowledge graphs in the field of aquatic disease prevention and control. Therefore, the construction and completion of knowledge graphs in the field of aquatic disease prevention and control urgently requires a dedicated technical architecture. This framework needs to possess the following key capabilities: accurate semantic parsing, especially resolving domain-specific semantic ambiguities; deep injection of knowledge from the field of aquatic disease prevention and control, guiding the model architecture to capture key information; designing a temporal collaborative fusion mechanism to ensure that new knowledge can be efficiently and accurately integrated into the existing knowledge system; and the ability to discover implicit knowledge based on logical reasoning. Summary of the Invention

[0006] This invention addresses the aforementioned problems in knowledge graph completion in the field of aquatic disease prevention and control by proposing a dynamic knowledge graph completion method based on DeepSeek. This method designs a knowledge graph completion framework called LongNSR-DeepSeek. First, to address the phenomenon of "different strategies for the same disease," an entity classification and labeling strategy based on the hazardous object is designed. By introducing a DeepSeek model optimized for vertical domain prompt templates and combining a global-local two-stage collaborative enhancement retrieval mechanism, the problem of context dependency failure in long texts is effectively alleviated, improving the accuracy and consistency of semantic retrieval. Second, to enhance the understanding of the deep semantic relationships between triples exhibiting "different strategies for the same disease," a differentiable logic rule enhancement module is adopted to achieve deep knowledge graph completion and reasoning optimization.

[0007] The technical solution of this invention is as follows:

[0008] The first aspect of this invention proposes a method for dynamically completing a knowledge graph of aquatic disease prevention and control based on DeepSeek, comprising the following steps:

[0009] S1: Construct a dataset for the prevention and control of aquatic diseases, including basic structured data and incremental data from academic papers;

[0010] S2: An entity classification and labeling mechanism based on hazardous objects is adopted, an attribute deconstruction strategy is introduced, and an extraction model is constructed by combining predefined schema constraints and the DeepSeek model. The extraction model is then used to extract triples from basic structured data to construct a basic knowledge graph.

[0011] S3: Using a global-local two-stage collaborative retrieval architecture, the DeepSeek model with vertical domain prompt template optimization is called to perform semantic enhancement retrieval on the incremental data of the academic papers, obtain high-quality text fragments, and input them into the extraction model in S2 for joint extraction of entities and relations, and finally form a new paper knowledge graph containing harmful object tags.

[0012] S4: Employ a time-assisted entity disambiguation strategy to merge the basic knowledge graph with the new paper knowledge graph into a high-quality knowledge graph with unified structure and consistent semantics;

[0013] S5: Input the high-quality knowledge graph into the differentiable logic rule enhancement module based on neural-symbolic fusion for further reasoning to obtain an optimized knowledge graph that has been enhanced with deep completion and logical consistency.

[0014] Furthermore, the process of constructing the basic structured data is as follows:

[0015] Define an initial data template with the disease name as the core entity and containing 8 key attributes; the key attributes include disease type, pathogen / cause, affected population, main symptoms, prevalence, treatment methods, control measures, and prevention methods;

[0016] Based on the initial data template, aquatic disease data from different sources are searched using the names of aquatic diseases as keywords, and then cleaned, deduplicated, and standardized to obtain basic structured data.

[0017] Furthermore, the process of constructing the incremental data of academic papers is as follows:

[0018] Search academic databases using the names of aquatic diseases as keywords to collect relevant academic papers;

[0019] To improve the accuracy and efficiency of subsequent knowledge extraction, the content of academic papers in PDF format is preprocessed, including: using prompt word engineering combined with the DeepSeek large language model to automatically identify and delete irrelevant data such as authors, affiliations, and references, while retaining the main text containing core academic arguments, thus obtaining incremental data of academic papers.

[0020] Furthermore, in step S2, the process of constructing the basic knowledge graph includes:

[0021] By combining schema constraints, entity classification and labeling mechanisms based on hazard objects, and attribute deconstruction strategies with the DeepSeek model, an extraction model is obtained.

[0022] Based on the extraction model, triples are extracted from the basic structured data to form a basic knowledge graph.

[0023] Furthermore, the schema constraints are defined based on the initial data template, including entity type, attribute type, triple composition format, and extraction processing rules; wherein,

[0024] The triple is represented in the form of (head entity, relation, tail entity), where the head entity is further refined into a structure of (head entity name, head entity label) based on the entity classification and labeling mechanism of the hazard object, so as to enhance the semantic level of the entity description and the accuracy of classification.

[0025] The entity types are: disease name, disease type, pathogen / cause, affected population, main symptoms, prevalence, treatment methods, control measures, and prevention methods;

[0026] The attribute types are: disease type, pathogen / cause, affected population, main symptoms, prevalence, treatment methods, control measures, and prevention methods;

[0027] The extraction processing rules are as follows: when the length of the tail entity text exceeds a preset threshold, its boundary is determined based on the context semantics and it is truncated; when a head entity and multiple tail entities have the same attribute, it is divided into multiple independent triples.

[0028] Furthermore, the attribute destructuring strategy includes:

[0029] The "epidemic situation" attribute is deconstructed and subdivided into three independent sub-attributes: epidemic time, epidemic location, and epidemic temperature, in order to separate spatiotemporal and environmental factor information.

[0030] The "main symptom" attribute is deconstructed into two sub-attributes: surface symptoms and internal symptoms, in order to distinguish the manifestation characteristics of the disease in different physiological locations.

[0031] Furthermore, in step S3, the process of constructing the new paper's knowledge graph includes:

[0032] Design a prompt template for the field of aquatic disease prevention and control. The template clarifies the entity types to be extracted, the extraction steps, and the output format, and provides multiple real aquatic disease cases to enhance the model's scenario adaptability. Use this prompt template to enhance the DeepSeek model.

[0033] Construct a global-local two-stage collaborative retrieval architecture: First, use the DeepSeek model optimized with the prompt template to perform a global retrieval, quickly locating macro-level documents or chapters related to the target disease from incremental academic paper data; then, based on the global retrieval results, initiate a local retrieval, focusing on fine-grained sentences or text fragments within the relevant documents or chapters to obtain high-quality text fragments.

[0034] The extraction model constructed in step S2 is used to extract entities and relations from the high-quality text fragments, resulting in triples with the same format as in step S2. These triples constitute the knowledge graph of the new paper.

[0035] Furthermore, the specific implementation of the global-local two-level collaborative retrieval architecture for obtaining high-quality text fragments is as follows:

[0036] The incremental data of academic papers and queries are fed into the DeepSeek model optimized with vertical domain prompt templates to obtain global search results;

[0037] Based on the global search results, a local search is initiated. The global search results and the query are fed into a cross-attention mechanism. The global search results are converted into value vectors and key vectors for the attention module, and the query is converted into a query vector for the attention module. The transpose of the query and key is calculated and normalized by Softmax to obtain the attention weight. This weight is then weighted with the value to obtain the association feature Z, which captures the semantic association between the "question" and the "global answer" and finds the part of the global answer that is most relevant to the question. Then, the association feature Z is processed by a fully connected layer and the gating coefficient is obtained by Sigmoid activation. Then, using the gating formula: Dynamic fusion of problem vectors With context vector This strengthens the relevance of semantic representation, ultimately resulting in an optimized local query that incorporates global information.

[0038] The local query, which incorporates global information, is then fed back into the DeepSeek model in S3.1, which is optimized using the vertical domain prompt template, to obtain the final retrieval result, namely, a high-quality text fragment.

[0039] Furthermore, in step S4, the method for obtaining a high-quality knowledge graph is as follows:

[0040] First, all triples are assigned timestamps to enhance disambiguation features: the basic knowledge graph is marked with the base time, and the knowledge graph of new papers is marked with their publication year;

[0041] Subsequently, entity alignment operations are performed on the head entities in the basic knowledge graph and the new paper knowledge graph to accurately identify and associate knowledge variants from different sources. The basic knowledge graph and the new paper knowledge graph are then merged to obtain a high-quality knowledge graph.

[0042] Furthermore, in step S5, the processing procedure of the differentiable logic rule enhancement module based on neural-symbolic fusion includes:

[0043] The high-quality knowledge graph is input into a neural knowledge graph embedding model (such as ComplEx) to learn vector representations of entities and relations, capturing latent semantic associations in the data, and obtaining the scores predicted by the fusion neural model. ;

[0044] Meanwhile, domain expert knowledge (such as "same disease + same target of harm → consistent prevention and control measures" and "prevention and control strategies at different times have temporal exclusivity" in the "same disease, different strategies" framework) is formalized into soft constraints of differentiable first-order logic rules. These rules are obtained by observing data and discussing with aquaculture experts multiple times.

[0045] These first-order logic rules are embedded into a differentiable computation graph through a logic tensor layer, and the rule reasoning score of each rule on the candidate triples is calculated. ;

[0046] Will and Adaptive weighted fusion is performed to obtain the final probability of the triplet being true. :

[0047]

[0048] in, It is a learnable balancing parameter used to dynamically adjust the contribution ratio of data-driven and rule-driven approaches.

[0049] Score predicted by fusion neural model Score based on rule reasoning Calculate a final probability of success for each candidate triplet. The range of values ​​is New triples with high confidence (range 0.7~1.0, adjusted according to actual scenario) are selected. Finally, these new relations are merged with the original high-quality knowledge graph to output an optimized knowledge graph that includes new inference triples, is semantically consistent, and has a clear temporal sequence after completion.

[0050] Furthermore, during the training process of the differentiable logic rule enhancement module, by jointly optimizing the data loss and rule satisfaction loss, the model can satisfy domain logic constraints as much as possible while fitting the observed data, thereby achieving accurate knowledge graph completion, implicit relation reasoning, and enhanced logical consistency; the joint loss function is as follows:

[0051]

[0052] in, This is the standard binary cross-entropy loss, used to fit the training data; This is the rule satisfaction loss, used to encourage the model to output prediction results that conform to preset logical constraints; This is the hyperparameter for rule strength.

[0053] A second aspect of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor; when the processor executes the computer program, the electronic device executes the DeepSeek-based dynamic completion method for aquatic disease prevention and control knowledge graph.

[0054] A third aspect of the present invention provides a storage medium comprising a computer program that, when run on an electronic device, causes the electronic device to execute the DeepSeek-based method for dynamically completing aquatic disease prevention and control knowledge graphs.

[0055] The advantages of this invention compared to existing technologies are as follows:

[0056] 1) Effectively resolve semantic ambiguity caused by "different solutions for the same problem".

[0057] By introducing an entity classification and labeling mechanism based on the hazard object, disease entities are refined into composite semantic units of "(disease, hazard object)," which fundamentally distinguishes the differences in prevention and control strategies for different hosts and at different times. This avoids the erroneous associations caused by semantic confusion in traditional methods and significantly improves the semantic clarity and logical consistency of the knowledge graph.

[0058] 2) Significantly improves the accuracy and completeness of knowledge extraction from long texts.

[0059] To address the characteristics of academic papers on aquatic diseases being lengthy and densely packed with technical terms, a global-local two-stage collaborative enhancement retrieval strategy is proposed. Combined with a DeepSeek model optimized by vertical domain prompt templates, this strategy effectively alleviates the context dependency failure problem in processing long texts in large models and improves the recall rate and extraction accuracy of key information.

[0060] 3) Achieve highly reliable reasoning completion driven by domain knowledge

[0061] The Differentiable Logic Rule Enhancement Module (NSR-KGC) is designed to integrate expert knowledge into neural representation learning in the form of symbolic logic, achieving synergistic optimization of data-driven and rule-driven approaches. This mechanism not only improves completion accuracy but also enhances model interpretability, making it particularly suitable for deep reasoning in complex scenarios such as "different solutions for the same problem."

[0062] 4) Supports dynamic time-series knowledge fusion and updating

[0063] By using time-series labels and entity disambiguation strategies, the organic integration of new and old knowledge is achieved, ensuring that the knowledge graph can integrate the latest research results in a timely manner and maintain its timeliness. This overcomes the limitations of traditional static knowledge graphs, which are slow to update and difficult to adapt to the evolution of disease prevention and control strategies.

[0064] 5) Provide reliable technical support for intelligent prevention and control of aquatic diseases.

[0065] Through precise knowledge extraction, semantic disambiguation, and logical reasoning, this method can provide targeted prevention and control solutions for different aquatic animals, support intelligent disease diagnosis, precise medication, and risk warning, and help improve the scientific and intelligent level of aquaculture management.

[0066] 6) Promote methodological innovation in the construction of knowledge graphs in vertical domains

[0067] This invention proposes a complete technical framework from knowledge extraction and fusion to reasoning, providing a systematic solution for the construction and dynamic completion of knowledge graphs in vertical fields such as agriculture and medicine, and has strong reference value. Attached Figure Description

[0068] Figure 1 This is a schematic diagram of the two-stage collaborative retrieval architecture of the present invention.

[0069] Figure 2 This is an example before the completion of a certain disease in the knowledge graph of this invention.

[0070] Figure 3 This is an example of a disease completed in the knowledge graph of this invention. Detailed Implementation

[0071] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0072] This invention provides a method for dynamically completing a knowledge graph for aquatic disease prevention and control based on DeepSeek, comprising the following steps:

[0073] S1: Construct a dataset for aquatic disease prevention and control, including basic structured data and incremental data from academic papers; the specific process is as follows:

[0074] S1.1: Define an initial data template with the disease name as the core entity and containing 8 key attributes; the key attributes include disease type, pathogen / cause, affected population, main symptoms, prevalence, treatment methods, control measures, and prevention methods.

[0075] S1.2: Based on the initial data template, search for aquatic disease data from different sources using the names of aquatic diseases as keywords, and perform cleaning, deduplication, and standardization processing to obtain basic structured data.

[0076] S1.3: Search academic databases using the names of aquatic diseases as keywords to collect relevant academic papers; to improve the accuracy and efficiency of subsequent knowledge extraction, preprocess the content of the PDF academic papers, including: using prompt word engineering combined with the DeepSeek large language model to automatically identify and delete irrelevant data such as authors, institutions, and references, while retaining the main text containing core academic arguments, thus obtaining incremental data of academic papers.

[0077] S2: An entity classification and labeling mechanism based on hazardous objects is adopted, an attribute deconstruction strategy is introduced, and an extraction model is constructed by combining predefined schema constraints and the DeepSeek model. This extraction model is then used to extract triples from basic structured data, thereby constructing a basic knowledge graph. The specific steps are as follows:

[0078] S2.1: Define schema constraints based on the defined initial data template. These schema constraints include entity type, attribute type, triplet composition format, and extraction processing rules; wherein,

[0079] The triples are represented in the form of (head entity, relation, tail entity), where the head entity is further refined into a structure of (head entity name, head entity label) based on the entity classification and labeling mechanism of the harmful object, to enhance the semantic hierarchy and classification accuracy of the entity description. For example: head entity (gill rot, grass carp), relation (internal symptoms), tail entity (increased mucus);

[0080] The entity types are: disease name, disease type, pathogen / cause, affected population, main symptoms, prevalence, treatment methods, control measures, and prevention methods;

[0081] The attribute types are: disease type, pathogen / cause, affected population, main symptoms, prevalence, treatment methods, control measures, and prevention methods;

[0082] The extraction processing rules are as follows: when the length of the tail entity text exceeds a preset threshold, its boundary is determined based on the context semantics and it is truncated; when a head entity and multiple tail entities have the same attribute, it is divided into multiple independent triples.

[0083] S2.2: To further improve the granularity and accuracy of knowledge, an attribute deconstruction strategy is implemented during the extraction process. This strategy can parse compound statements into multiple precise and independent triples, including:

[0084] The "epidemic situation" attribute is deconstructed into three independent sub-attributes: epidemic time, epidemic location, and epidemic temperature, to separate spatiotemporal and environmental factors. For example: the epidemic situation is prevalent from March to June, with water temperatures of 14-18℃, and it is more common in juvenile fish weighing 10-20g, but can also affect adult fish weighing 50-60g. This disease occurs annually in fish farms in Tokushima and Waka prefectures, Japan, the main production areas of sweetfish. After deconstruction, it is divided into: epidemic time from March to June, epidemic location in fish farms in Tokushima and Waka prefectures, Japan, and epidemic temperature of 14-18℃.

[0085] The "main symptom" attribute is deconstructed into two sub-attributes: surface symptoms and internal symptoms, to differentiate the disease's manifestations in different physiological locations. For example, the main symptom might be characteristic symptoms such as surface erosion, ulcers, and white spots on the liver, with focal necrosis of the liver visible on histopathological examination. In some surviving fish, a tumor may form centered around the mouth. After deconstruction, this can be divided into: surface symptoms are characteristic symptoms such as surface erosion, ulcers, and white spots on the liver; internal symptoms are focal necrosis of the liver, with a tumor forming centered around the mouth in some surviving fish.

[0086] S2.3: Combine the schema constraints and entity classification and labeling mechanism based on hazard objects described in S2.1, and the attribute deconstruction strategy described in S2.2 with the DeepSeek model to obtain the final extraction model.

[0087] S2.4: Based on the extraction model, triples are extracted from the basic structured data to form a basic knowledge graph.

[0088] S3: A global-local two-stage collaborative retrieval architecture is used to call the DeepSeek model with vertical domain prompt template optimization to perform semantic enhancement retrieval on the incremental data of academic papers, obtaining high-quality text fragments, which are then input into the extraction model for joint extraction of entities and relations, ultimately forming a new paper knowledge graph containing harmful object tags; specifically including:

[0089] S3.1: Design a prompt template for the field of aquatic disease prevention and control. The template clarifies the entity types to be extracted, the extraction steps and the output format, and provides multiple real aquatic disease cases to enhance the model's scenario adaptability. Use the prompt template to enhance the DeepSeek model.

[0090] S3.2: Construct a global-local two-stage collaborative retrieval architecture: First, perform a global retrieval to quickly locate macro-level documents or chapters related to the target disease from the incremental data of academic papers; then, based on the global retrieval results, initiate a local retrieval to focus on fine-grained sentences or text fragments within the relevant documents or chapters to obtain high-quality text fragments.

[0091] S3.2.1: Input the incremental data of academic papers and queries into the DeepSeek model optimized by the vertical domain prompt template in step S3.1 to obtain the global search results;

[0092] S3.2.1: Based on the global search results, a local search is initiated. The global search results and the query are fed into a cross-attention mechanism. The global search results are converted into the value vector and key vector of the attention module, and the query is converted into the query vector of the attention module. The transpose of the query and the key is calculated and normalized by Softmax to obtain the attention weight. This weight is then weighted with the value to obtain the association feature Z, which captures the semantic association between the "question" and the "global answer" and finds the part of the global answer that is most relevant to the question. Then, the association feature Z is processed by a fully connected layer and the gating coefficient is obtained by Sigmoid activation. Then, using the gating formula: Dynamic fusion of problem vectors With context vector This strengthens the relevance of semantic representation, ultimately resulting in an optimized local query that incorporates global information.

[0093] S3.2.3: The local query that integrates global information is fed back into the DeepSeek model optimized by the vertical domain prompt template in step S3.1 to obtain the final retrieval result, i.e., high-quality text fragments.

[0094] S3.3: Based on the extraction model constructed in step S2.3, entity and relation extraction is performed on the high-quality text fragments to obtain triples with the same format as in step S2. These triples constitute the knowledge graph of the new paper.

[0095] S4: Employing a time-assisted entity disambiguation strategy, the basic knowledge graph and the new paper's knowledge graph are merged into a high-quality knowledge graph with unified structure and consistent semantics; the specific process is as follows:

[0096] First, all triples are assigned timestamps to enhance disambiguation features: the basic knowledge graph is marked with the base time, and the knowledge graph of new papers is marked with their publication year;

[0097] Subsequently, entity alignment operations are performed on the head entities in the basic knowledge graph and the new paper knowledge graph to accurately identify and associate knowledge variants from different sources. The basic knowledge graph and the new paper knowledge graph are then merged to obtain a high-quality knowledge graph.

[0098] S5: Input the knowledge graph obtained in S4 into the differentiable logic rule enhancement module based on neural-symbolic fusion for further reasoning, resulting in an optimized knowledge graph enhanced with deep completion and logical consistency; the specific process is as follows:

[0099] The high-quality knowledge graph is input into a neural knowledge graph embedding model (such as ComplEx) to learn vector representations of entities and relations, capturing latent semantic associations in the data, and obtaining the scores predicted by the fusion neural model. ;

[0100] Meanwhile, domain expert knowledge (such as "same disease + same target → consistent prevention and control measures" and "prevention and control strategies at different times have temporal exclusivity" in the "different strategies for the same disease" framework) is formalized into soft constraints of differentiable first-order logic rules. These rules are obtained through observing data and discussing with aquaculture experts multiple times. For example, the rules can be expressed as:

[0101]

[0102] These first-order logic rules are embedded into a differentiable computation graph through a logic tensor layer, and the rule reasoning score of each rule on the candidate triples is calculated. ;

[0103] Will and Adaptive weighted fusion is performed to obtain the final probability of the triplet being true. :

[0104]

[0105] in, It is a learnable balancing parameter used to dynamically adjust the contribution ratio of data-driven and rule-driven approaches.

[0106] During the training of the differentiable logic rule enhancement module, by jointly optimizing the data loss and rule satisfaction loss, the model can satisfy domain logic constraints as much as possible while fitting the observed data, thereby achieving accurate knowledge graph completion, implicit relation reasoning, and enhanced logical consistency. The joint loss function is as follows:

[0107]

[0108] in, This is the standard binary cross-entropy loss, used to fit the training data; This is the rule satisfaction loss, used to encourage the model to output prediction results that conform to preset logical constraints; This is the hyperparameter for rule strength.

[0109] Score predicted by fusion neural model Score based on rule reasoning Calculate a final probability of success for each candidate triplet. The range of values ​​is New triples with high confidence (range 0.7~1.0) are selected, and these new relations are finally merged with the original high-quality knowledge graph to output an optimized knowledge graph that includes new inference triples, is semantically consistent, and has a clear temporal sequence after completion.

[0110] 1. Verification Experiment

[0111] The operating system used was Ubuntu 22.04, the CPU was an Intel Xeon Platinum 8470Q (20 vCPUs), the GPU was an NVIDIA RTX 4090 (24 GB), and the deep learning framework was PyTorch 2.5.1. The experiment consisted of 10 training rounds using the Adam optimizer with an initial learning rate of 0.01, supplemented by a warm-up learning rate strategy (3 warm-up rounds, an initial learning rate of 0.1, and a momentum of 0.8). The training batch size was set to 256, and the embedding dimension for both entities and relations was 100. During training, a rule strength coefficient was introduced. =0.1 was used to balance neural loss and rule loss, ensuring that the model effectively incorporates domain logic constraints on a data-driven basis. Experiments were conducted using link prediction metrics (including MR, MRR, and Hits@N) widely used in knowledge graph completion tasks for evaluation. All results were obtained on tail entity prediction tasks, validating the effectiveness of the proposed method.

[0112] To systematically evaluate the scale and quality of the aquatic disease prevention and control knowledge graph constructed in this invention, a comprehensive statistical analysis was conducted on the fused knowledge graph, and the results are shown in Table 1. This knowledge graph is composed of two core parts: sub-graph A, constructed based on the original basic structured data, and sub-graph B, extracted from incremental data from academic papers. As shown in Table 1, sub-graph A contains 2120 entities and 9928 triples, covering basic and classic disease knowledge in the field; sub-graph B, through fine-grained extraction and object annotation, contributed 4801 entities and 3503 triples, significantly expanding the frontier and coverage of the knowledge graph. After temporal fusion and entity disambiguation, the final unified aquatic disease prevention and control knowledge graph contains 533 diseases, 6619 entities, and 13431 triples, with 70% of the triples exhibiting "different strategies for the same disease." The entity size has increased by 212.2% compared to the basic subgraph A, and the number of triples has increased by 35.3%. This significant expansion fully demonstrates the huge incremental contribution of new paper data to the knowledge system, effectively enhancing the completeness and timeliness of the knowledge graph.

[0113] Table 1. Knowledge Graph Statistics

[0114]

[0115] In particular, the "entity classification and labeling based on the harmful object" mechanism proposed in this invention plays a core role in the knowledge fusion process, effectively solving the serious relational ambiguity problem caused by the phenomenon of "different treatments for the same disease". This mechanism refines the original single "disease-relationship-tail entity" triplet into a semantically refined "(disease, harmful object)-relationship-tail entity" by attaching specific aquatic animal tags (such as "grass carp" and "whiteleg shrimp") to disease entities. For example, for the entity "gill rot", traditional methods cannot distinguish its differences in different hosts, while this method successfully constructs two semantically clear and non-interfering triplets: "(gill rot, grass carp)-prevention measures-chlorine dioxide disinfection" and "(gill rot, crucian carp)-prevention measures-traditional Chinese medicine bait". This fine-grained entity representation fundamentally improves the semantic clarity and logical consistency of the knowledge graph, laying a solid foundation for subsequent high-precision knowledge reasoning and completion.

[0116] 2. Comparative Experiment

[0117] (1) Based on vertical domain datasets

[0118] To comprehensively evaluate the performance of the LongNSR-DeepSeek knowledge graph completion framework of this invention in vertical domain knowledge graph completion tasks, four representative open-domain KGC methods (DKRL, ConMask, Cmplx-OWE, and RST) were selected as baseline models and comparative experiments were conducted on the aquatic disease prevention and control knowledge graph. The results are shown in Table 2.

[0119] Table 2 Experimental results of KGC models in each open domain

[0120]

[0121] As shown in Table 2, our method significantly outperforms the baseline model across all evaluation metrics. Specifically, the MR value is only 48.60, approximately 80.8% lower than the best baseline model RST (252.5), indicating that our method can more accurately rank the correct answers at the top. The MRR reaches 0.725, a 98.6% improvement over RST, reflecting a significant enhancement in the overall predictive stability of the model. Hits@1, Hits@3, and Hits@10 reach 0.635, 0.805, and 0.865 respectively, representing improvements of 135.2%, 98.8%, and 61.7% over RST. The significant improvement in Hits@1 is particularly crucial, indicating that our framework has a stronger ability to provide the "first-choice correct answer" in practical applications.

[0122] Meanwhile, the baseline model has three limitations in vertical domains. First, there is a mismatch in the granularity of entity semantics. While methods such as DKRL and Cmplx-OWE perform well in open domains, they fail to adapt to the fine-grained semantic distinctions required for "different treatments for the same disease" in the field of aquatic diseases. For example, for the entity "gill rot," open domain models cannot distinguish the differences in prevention and control in different hosts (such as "grass carp" and "crucian carp"), while this framework, through the "harmful object labeling" mechanism, refines the disease entity into a composite unit of "(disease, harmful object)," effectively solving this kind of semantic ambiguity problem. Second, the quality of knowledge extraction is insufficient. When processing long texts of academic papers on aquatic diseases, methods such as ConMask and RST suffer from decreased extraction accuracy due to the failure of context dependencies. For example, when processing long sentences such as "grass carp are prone to gill rot when the water temperature in freshwater ponds is 25-30℃ in summer, and chlorine dioxide can be used for disinfection," ConMask is prone to mistakenly associating "water temperature 25-30℃" with the "treatment method" attribute. This framework utilizes LongLGN's "global-local two-stage collaborative retrieval" mechanism to first locate relevant chapters on "epidemic status" and then focus on key statements, effectively filtering out noise interference. Thirdly, the reasoning logic lacks domain constraints. Baseline models generally rely on purely data-driven neural reasoning, failing to introduce domain prior logic, making it difficult to handle complex scenarios such as "different treatments for the same disease." For example, open-domain models are prone to erroneous associations regarding the differences in prevention and control measures for "vibriovirus" on different hosts. This framework, through the NSR-KGC module, transforms domain rules (such as "same disease + same harmful target → consistent treatment plan") into differentiable constraints, significantly improving reasoning accuracy.

[0123] Figure 2 and Figure 3 The table shows the visualized knowledge graphs before and after completion using the LongNSR-DeepSeek framework of this invention. As can be seen from Table 2, this method, through its dual-core design of "high-quality knowledge input + domain logic reasoning," achieves a high degree of compatibility with KGC tasks in the aquatic disease domain. From a data perspective, it integrates basic structured data with incremental data from academic papers to construct a full-spectrum knowledge system covering classic disease knowledge and cutting-edge prevention and control technologies. From an extraction perspective, the LongLGN module addresses the issues of long text noise and fine-grained ambiguity in vertical domains through global-local retrieval and hazard object annotation. From a reasoning perspective, the NSR-KGC module transforms domain expert knowledge into differentiable rules through neural-symbolic fusion, ensuring that model reasoning relies on data statistics while conforming to the actual application logic of aquatic diseases. Experimental results demonstrate that this framework has significant advantages in vertical domain KGC tasks, laying a technical foundation for the intelligent application of aquatic disease prevention and control knowledge.

[0124] (2) Based on general domain datasets

[0125] To further verify the generalization ability of the proposed LongNSR-DeepSeek model in general open domain scenarios, a comparative experiment was conducted using the classic public dataset DBPedia50k in the knowledge graph completion field. This dataset contains 49,900 entities within the knowledge graph, 654 relations, and 32,388 training triples. The data format is a standard (head entity-relation-tail entity) structure with no domain-specific labels, and can be used to test the model's basic completion performance in scenarios where different solutions are used for the same problem. The specific experimental design process is as follows:

[0126] Dataset preprocessing: Following the official DBPedia50k dataset partitioning (32,388 training records, 399 validation records, and 2,001 test records), the "Filtered" evaluation setting was adopted (excluding triples already existing in the knowledge graph to avoid duplicate scoring). Since this dataset lacks entity label information, the "hazard object classification and annotation" module was not required for the model; only the core Schema-optimized knowledge extraction (SDeepSeek), global-local two-order collaborative retrieval (LongLGN), and neural-symbolic logic reasoning (NSR-KGC) functions were retained to adapt to general open-domain data characteristics.

[0127] Model Selection for Comparison: Five representative open-domain knowledge graph completion models were selected as baselines, covering traditional descriptive, structural, and recent advanced models to ensure fairness in the comparison. ConMask: A descriptive model based on CNN and relation-dependent attention, adept at generating embeddings using entity text descriptions; Cmplx-OWE: A model combining complex space embeddings and open-world entity transformation functions to optimize open-domain entity completion; RST: A logic-enhanced model based on relation structure tensors, capturing semantic associations through tensor decomposition; MIA: A multi-view attention fusion model that integrates entity structure and text features to improve completion accuracy; IKGE: An inductive model based on attention feature aggregation, supporting multi-hop neighbor information fusion.

[0128] Table 3 Experimental results of KGC models in each open domain

[0129]

[0130] As shown in Table 3, the experimental results demonstrate that, compared to existing open-domain KGC methods, the LongNSR-DeepSeek model proposed in this invention outperforms most other metrics on the DBPedia50k dataset. Specifically, the average ranking metric MR (16.8) is significantly lower than other models, decreasing by 11.6% compared to the second-best MIA (19), indicating that the model ranks correct tail entities higher. The comprehensive score metric MRR (0.67) is higher than MIA (0.63), IKGE (0.61), and Cmplx-OWE (0.603), reflecting stronger overall prediction stability. In terms of hit rate, Hit@3 (0.73) and Hit@10 (0.86) are the best, with Hit@10 showing a 7.5% improvement over RST (0.80) and a 3.6% improvement over MIA (0.83). Although it is slightly lower than RST (0.56) in the Hit@1 metric, LongNSR-DeepSeek (0.55) is still better than MIA (0.50) and Cmplx-OWE (0.52), and maintains a comprehensive advantage in overall performance, proving that the LongNSR-DeepSeek model of this invention has higher accuracy and robustness in candidate entity ranking.

[0131] In the unlabeled DBPedia50k scenario, LongLGN's global-local two-stage collaborative retrieval mechanism can still efficiently locate structured information related to entities (such as entity attributes and relationships), reducing noisy triple inputs and providing a high-quality data foundation for subsequent inference. This is the key to the significant reduction in the model's MR. Compared to RST, which relies solely on tensor decomposition to capture relational structures, NSR-KGC further optimizes the embedding learning process by incorporating general domain logic rules (such as "entity type consistency constraints" and "relation transitivity soft constraints"), enabling MRR and most hit rate metrics to surpass purely data-driven models such as MIA. Although DBPedia50k lacks domain labels, SDeepSeek's schema fine-grained reconstruction capabilities (such as semantic decomposition of "attribute-value" relationships) can still improve the purity of triple extraction, avoiding prediction bias caused by mixed attribute semantics in traditional models.

[0132] In summary, the LongNSR-DeepSeek model of this invention performs excellently in both the aquatic disease domain (labeled, "different strategies for the same disease" scenario) and the DBPedia50k general domain (unlabeled, standard triplet scenario), demonstrating that its core mechanism (retrieval-enhanced knowledge extraction + neural-symbolic reasoning) has cross-scenario adaptability. That is, it does not rely on domain-specific labels; by optimizing the quality of knowledge input and introducing general logical constraints, it can achieve high-performance completion in open-domain tasks, solving the problems of "overfitting domain features and poor generalization" in traditional domain models. Experimental results show that this model is not only effective in the "different strategies for the same disease" scenario in the aquatic disease domain, but also exhibits excellent performance and strong generalization ability in the standard triplet completion task in the general open domain, providing strong support for its application in a wider range of knowledge graph completion scenarios.

[0133] 3. Ablation test

[0134] As shown in Table 4, the DeepSeek+LongLGN (denoted as Model-A), using the DeepSeek model with schema constraints as the core for entity relation extraction and combining it with LongLGN's global-local two-stage collaborative retrieval mechanism, achieved benchmark performance (MR=65.80, MRR=0.675, Hits@10=0.765) in the tail entity prediction task. This result shows that the refined schema reconstruction (such as deconstructing "popularity" into "popularity time / location / temperature") effectively reduces the semantic clutter of attributes and improves the purity of triple extraction; at the same time, LongLGN's two-stage retrieval mechanism (first globally locating relevant chapters, then locally focusing on key sentences) alleviates the context dependency failure problem caused by long texts.

[0135] Table 4 Ablation Experiment

[0136]

[0137] By comparing the performance of Model-A (SDeepSeek+LongLGN) and Model-B (SDeepSeek+NSR-KGC) in the ablation experiment in Table 4, it can be seen that when the NSR-KGC module is introduced alone, the model's median rank (MR) improves to 72.40 (the lower the value, the better the ranking), but the MRR, Hits@3, and Hits@10 decrease slightly (to 0.648, 0.680, and 0.755, respectively). This phenomenon indicates that the NSR-KGC module significantly enhances the reasoning ability for ambiguous relationships such as "different treatments for the same disease" by transforming domain logic rules (e.g., "same disease + same harmful object → consistent treatment plan") into differentiable constraints (e.g., distinguishing different treatment plans for "grass carp gill rot" and "crucian carp gill rot"), thereby improving the median rank (MR) of correct triples. However, because this module does not optimize the quality of long text knowledge extraction, noisy triples still exist in the original data (e.g., the harmful object is unclear due to context dependency failure), affecting the overall prediction stability (MRR). This indicates that the core value of the NSR-KGC module lies in improving inference accuracy, but its performance depends on high-quality knowledge input.

[0138] Model-C (LongNSR-DeepSeek) integrates the high-quality knowledge extraction of LongLGN with the logical reasoning ability of NSR-KGC, achieving optimal performance across all metrics: MR is reduced to 48.60 (a 26.1% decrease compared to Model-A), MRR is increased to 0.725 (a 7.4% increase compared to Model-A), and Hits@1, Hits@3, and Hits@10 reach 0.635, 0.805, and 0.865 respectively (12.4%, 14.2%, and 13.1% increases compared to Model-A). This result fully verifies the synergistic effect of the technical framework of this invention: LongLGN reduces knowledge extraction noise through retrieval enhancement, providing NSR-KGC with "clean" triple input, enabling logical rules to more accurately constrain the reasoning process; for the typical "different treatments for the same disease" problem in the field of aquatic diseases (such as different control measures for "vibrio disease" in "Litopenaeus vannamei" and "turbot"), LongLGN's "hazard object classification labeling" provides key semantic features for NSR-KGC's rule reasoning (such as "(vibrio disease, Litopenaeus vannamei) - control measures - chlorine dioxide disinfection"), enabling the model to accurately distinguish entity semantic differences. The combination of LongLGN's extracted paper data and NSR-KGC rules enables the model to dynamically update knowledge, further improving the timeliness and accuracy of completion.

[0139] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the claims of the present invention.

Claims

1. A method for dynamically completing a knowledge graph of aquatic disease prevention and control based on DeepSeek, characterized in that, Includes the following steps: S1: Construct a dataset for the prevention and control of aquatic diseases, including basic structured data and incremental data from academic papers; S2: An entity classification and labeling mechanism based on hazardous objects is adopted, an attribute deconstruction strategy is introduced, and an extraction model is constructed by combining predefined schema constraints and the DeepSeek model. The extraction model is then used to extract triples from basic structured data to construct a basic knowledge graph. S3: Using a global-local two-stage collaborative retrieval architecture, the DeepSeek model with vertical domain prompt template optimization is called to perform semantic enhancement retrieval on the incremental data of the academic papers, obtain high-quality text fragments, and input them into the extraction model in S2 for joint extraction of entities and relations, and finally form a new paper knowledge graph containing harmful object tags. S4: Employ a time-assisted entity disambiguation strategy to merge the basic knowledge graph with the new paper knowledge graph into a high-quality knowledge graph with unified structure and consistent semantics; S5: Input the high-quality knowledge graph into the differentiable logic rule enhancement module based on neural-symbolic fusion for further reasoning to obtain an optimized knowledge graph that has been enhanced with deep completion and logical consistency.

2. The method for dynamically completing aquatic disease prevention and control knowledge graph based on DeepSeek according to claim 1, characterized in that, The process of constructing the basic structured data is as follows: Define an initial data template with the disease name as the core entity and containing 8 key attributes; the key attributes include disease type, pathogen / cause, affected population, main symptoms, prevalence, treatment methods, control measures, and prevention methods; Based on the initial data template, aquatic disease data from different sources are searched using the names of aquatic diseases as keywords, and then cleaned, deduplicated, and standardized to obtain basic structured data.

3. The method for dynamically completing a knowledge graph of aquatic disease prevention and control based on DeepSeek according to claim 1, characterized in that, The process of constructing the incremental data of academic papers is as follows: Search academic databases using the names of aquatic diseases as keywords to collect relevant academic papers; The content of academic papers in PDF format is preprocessed, including: using prompt word engineering combined with DeepSeek large language model to automatically identify and delete irrelevant data including authors, affiliations, and references, while retaining the main text containing core academic arguments, thus obtaining incremental data of academic papers.

4. The method for dynamically completing a knowledge graph of aquatic disease prevention and control based on DeepSeek according to claim 1, characterized in that, In step S2, the process of constructing the basic knowledge graph includes: By combining schema constraints, entity classification and labeling mechanisms based on hazard objects, and attribute deconstruction strategies with the DeepSeek model, an extraction model is obtained. Based on the extraction model, triples are extracted from the basic structured data to form a basic knowledge graph.

5. The method for dynamically completing aquatic disease prevention and control knowledge graph based on DeepSeek according to claim 4, characterized in that, The schema constraints are defined based on the initial data template, including entity type, attribute type, triple composition format, and extraction processing rules; among which, The triple is represented in the form of (head entity, relation, tail entity), where the head entity is further refined into a structure of (head entity name, head entity label) based on the entity classification and labeling mechanism of the hazard object, so as to enhance the semantic level of the entity description and the accuracy of classification. The entity types are: disease name, disease type, pathogen / cause, affected population, main symptoms, prevalence, treatment methods, control measures, and prevention methods; The attribute types are: disease type, pathogen / cause, affected population, main symptoms, prevalence, treatment methods, control measures, and prevention methods; The extraction processing rules are as follows: when the length of the tail entity text exceeds a preset threshold, its boundary is determined based on the context semantics and it is truncated; when a head entity and multiple tail entities have the same attribute, it is divided into multiple independent triples.

6. The method for dynamically completing a knowledge graph of aquatic disease prevention and control based on DeepSeek according to claim 4 or 5, characterized in that, The attribute destructuring strategy includes: The "epidemic situation" attribute is deconstructed and subdivided into three independent sub-attributes: epidemic time, epidemic location, and epidemic temperature, in order to separate spatiotemporal and environmental factor information. The "main symptom" attribute is deconstructed into two sub-attributes: surface symptoms and internal symptoms, in order to distinguish the manifestation characteristics of the disease in different physiological locations.

7. The method for dynamically completing aquatic disease prevention and control knowledge graph based on DeepSeek according to claim 6, characterized in that, In step S3, the process of constructing the knowledge graph of the new paper includes: Design a prompt template for the field of aquatic disease prevention and control. The template clarifies the entity types to be extracted, the extraction steps, and the output format, and provides multiple real aquatic disease cases to enhance the model's scenario adaptability. Use this prompt template to enhance the DeepSeek model. Construct a global-local two-stage collaborative retrieval architecture: First, use the DeepSeek model optimized with the prompt template to perform a global retrieval, quickly locating macro-level documents or chapters related to the target disease from incremental academic paper data; then, based on the global retrieval results, initiate a local retrieval, focusing on fine-grained sentences or text fragments within the relevant documents or chapters to obtain high-quality text fragments. The extraction model constructed in step S2 is used to extract entities and relations from the high-quality text fragments, resulting in triples with the same format as in step S2. These triples constitute the knowledge graph of the new paper.

8. The method for dynamically completing a knowledge graph of aquatic disease prevention and control based on DeepSeek according to claim 7, characterized in that, The specific implementation of the global-local two-level collaborative retrieval architecture for obtaining high-quality text fragments is as follows: The incremental data of academic papers and queries are fed into the DeepSeek model optimized with vertical domain prompt templates to obtain global search results; Based on the global search results, a local search is initiated. The global search results and the query are fed into a cross-attention mechanism. The global search results are converted into value vectors and key vectors for the attention module, and the query is converted into a query vector for the attention module. The transpose of the query vector and the key vector is calculated and normalized using Softmax to obtain the attention weights. These weights are then weighted with the value vectors to obtain the association feature Z, which captures the semantic association between the "question" and the "global answer," identifying the part of the global answer most relevant to the question. This association feature Z is then processed through a fully connected layer and activated by a Sigmoid layer to obtain the gating coefficients. Then, using the gating formula: Dynamic fusion of problem vectors With context vector This strengthens the relevance of semantic representation, ultimately resulting in an optimized local query that incorporates global information. The local query, which incorporates global information, is then fed back into the DeepSeek model, which optimizes the vertical domain prompt template, to obtain the final retrieval result, namely, a high-quality text fragment.

9. The method for dynamically completing a knowledge graph of aquatic disease prevention and control based on DeepSeek according to claim 7, characterized in that, In step S4, the method for obtaining a high-quality knowledge graph is as follows: To enhance disambiguation features, all triples are assigned timestamps: the basic knowledge graph is marked with the baseline time, and the knowledge graph of new papers is marked with their publication year. Entity alignment is performed on the head entities in the basic knowledge graph and the new paper knowledge graph to accurately identify and associate knowledge variants from different sources. The basic knowledge graph and the new paper knowledge graph are then merged to obtain a high-quality knowledge graph.

10. The method for dynamically completing a knowledge graph of aquatic disease prevention and control based on DeepSeek according to claim 7, characterized in that, In step S5, the processing procedure of the differentiable logic rule enhancement module based on neural-symbolic fusion includes: The high-quality knowledge graph is input into the neural knowledge graph embedding model to learn vector representations of entities and relationships, capturing latent semantic associations in the data, and obtaining the scores predicted by the fusion neural model. ; Meanwhile, the domain expert knowledge of "different solutions for the same problem" is formalized into soft constraints of differentiable first-order logic rules, which are obtained by observing data and discussing with aquaculture experts multiple times. These first-order logic rules are embedded into a differentiable computation graph through a logic tensor layer, and the rule reasoning score of each rule on the candidate triples is calculated. ; Will and Adaptive weighted fusion is performed to obtain the final probability of the triplet being true. : in, These are learnable equilibrium parameters; The range of values ​​is ; based on New triples with high confidence are selected, and these new relations are finally merged with the original high-quality knowledge graph to output an optimized knowledge graph that includes new inference triples, is semantically consistent, and has a clear temporal sequence.

Images