Method for identifying risks of special articles of entry and exit at port based on knowledge graph construction and entity linking

Abstract

The application relates to a kind of port entry and exit special goods risk identification methods based on knowledge graph construction and entity linking, comprising: according to the risk identification demand of port entry and exit special goods in advance, constructing hierarchical risk factor knowledge graph;Obtain the declaration material of risk to be identified, carry out entity extraction from the declaration material, retrieve historical similar cases according to the extracted entity, build enhanced context containing declaration material and its historical similar cases, and carry out overall vectorization processing;Adopt three-granularity matching strategy: first, through Top-K screening, to match special goods subclass with coarse granularity;Then through threshold filtering, to match risk factor with medium granularity, finally, fine-grained matching risk feature node, to obtain the risk identification result of port entry and exit special goods.Compared with the prior art, the present application has the advantages of more precise risk feature expression, explicit risk knowledge audit, and realizes accurate matching from declaration material to knowledge graph node.

Description

Technical Field

[0001] This invention relates to the field of risk identification technology for goods, and in particular to a method for risk identification of special goods entering and leaving ports based on knowledge graph construction and entity linking. Background Technology

[0002] Special items entering and leaving ports refer to microorganisms, human tissues, biological products, blood and their products, etc., which require declaration and approval. These items pose potential biosafety risks, infectious disease transmission risks, and ecological security risks. Under the current regulatory system, risk assessment of special items relies heavily on manual review and expert experience. Reviewers need to extensively consult information from various sources, including laws and regulations, scientific literature, epidemic reports, and product manuals, to determine the risk status of specific items under specific risk dimensions.

[0003] For example, the invention with publication number CN121233800A discloses a method and system for rapid retrieval of special items at ports based on data processing. By collecting representational information such as images and text of special items, and considering the complexity of the modal features of special items, it applies unified representation technology of image and object information, similarity measurement method, feature association technology, and uses industry knowledge graph to establish a retrieval model for biosafety risk monitoring. This enables the matching of biosafety target features for non-trade personal carrying and delivery channels, and the identification of high-risk factors for multilingual special items.

[0004] This solution simply searches and matches based on the image, text, and source information of the current item, and has the following main shortcomings: (1) Risk characteristics are highly heterogeneous and difficult to represent using traditional rule bases. The sources of risk for different subcategories of goods vary greatly: the main risk of nucleic acid products lies in the biological activity and recombination potential of their genetic material; the main risk of blood products lies in the spread of human pathogens; and recombinant viral vector products involve the pathogenicity of the vector itself, the functional characteristics of the packaging elements, and the potential effects of the inserted gene. This highly differentiated risk structure makes it difficult for rules based on a uniform structure to represent the risk knowledge of specific goods.

[0005] (2) Existing knowledge organization methods cannot support fine-grained risk reasoning. Although there is a catalog-based item classification system, this system only describes the category and basic characteristics of items, fails to effectively express the relationship between items and risk factors, and cannot support evidence retrieval and risk assessment of specific items under specific risk dimensions.

[0006] (3) Existing feature matching methods are difficult to accurately associate the declaration text of items with risk factors. New special items often have new components and new risk factors, and there are a large number of abbreviations and synonyms with different forms. Using precise keywords or fixed word embedding vectors for matching often fails to hit the correct concepts in the knowledge graph. An entity linking method that integrates domain knowledge and dynamic context is needed.

[0007] (4) Shortcomings of existing large model fine-tuning or external data RAG solutions. Fine-tuning large models compresses domain knowledge into model weights, making it impossible to audit the basis of their judgments, and knowledge updates require retraining, which poses a risk of illusion; ordinary RAG solutions retrieve original text fragments, the risk dimension boundaries are vague, they cannot support the fine decomposition of entity features, and they do not have the ability to express the correlation of risk features. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of the existing technology, such as highly heterogeneous risk characteristics, difficulty in representing them with traditional rule bases, inability to support fine-grained risk reasoning, and difficulty in accurately associating item declaration texts with risk factors, and to provide a method for identifying the risk of special items entering and leaving ports based on knowledge graph construction and entity linking.

[0009] The objective of this invention can be achieved through the following technical solutions: A method for risk identification of special items entering and leaving ports based on knowledge graph construction and entity linking includes the following steps: Based on the risk identification requirements for special items entering and leaving the port, a hierarchical risk factor knowledge graph is constructed in advance, and the corresponding content source text is obtained. The hierarchical risk factor knowledge graph has a multi-layer structure, which includes a root node for special items, a sub-category node for special items, a risk factor node, a risk feature node, and a risk level criterion node. Relevant information on special items and risk factors is extracted from the content source text and mapped to the nodes of each layer of the hierarchical risk factor knowledge graph. Obtain the application materials for the risks to be identified, extract entities from the application materials, retrieve historical similar cases based on the extracted entities, construct an enhanced context containing the application materials and their historical similar cases, and perform overall vectorization processing; The vectorized enhanced context is matched with the information vectors of all special item subclass nodes respectively, and one or more special item subclass nodes with the highest similarity are obtained. The corresponding risk factor nodes are added to the candidate subclass set. Within the candidate subclass set, the vectorized enhanced context is matched with the information vectors of each risk factor node respectively. Based on the similarity calculation results, the candidate risk factor set is selected. After pairing the elements in the candidate subclass set and the candidate risk factor set one by one, multiple query combinations are formed. For each query combination, the existence of a corresponding risk feature node is queried. Based on the query results, the risk feature node set and its confidence level and risk level criterion node information are output as the risk identification result of special items entering and leaving the port.

[0010] Furthermore, the root node for special items represents the general category of special items for entry and exit; The special item subclass node is a child node of the corresponding special item root node, and includes node identifier, classification definition, basic item feature description and inherent risk level attribute; The risk factor nodes include pathogenicity, toxicity, infectivity, drug resistance, environmental persistence, host range, and ecological safety risks. The risk feature node is generated by combining special item sub-category nodes and risk factor nodes based on expert knowledge and historical assessment records; The risk level criterion nodes include a set of keywords, textual descriptions, supporting documents, expert names, and risk level labels.

[0011] Furthermore, relevant information on special items and risk factors is extracted from the source text of the content, including: The source text of the content is subjected to entity recognition, relation extraction, and evidence-conclusion pair extraction. Entity recognition is used to obtain the item name, component keywords, and risk terms. Relationship extraction is used to identify the association between items and risk factors and the basis for judging the effectiveness of the combination. Evidence-conclusion pairs are used to extract the keywords, risk level labels, literature sources, and expert signatures on which each risk judgment is based.

[0012] Furthermore, the information extracted from the source text is mapped to the nodes of each layer of the hierarchical risk factor knowledge graph, including: Based on the entity recognition results, supplement the special item subclass nodes with classification definitions, basic item feature descriptions, and inherent risk level attributes; supplement the risk factor nodes with the scope of application for each risk factor; Based on the relationship extraction results, the effectiveness of the combination between special item subclass nodes and risk factor nodes is confirmed, and risk feature nodes are generated. Based on the evidence-conclusion analysis, supplement the risk level criterion nodes with the required keyword set, text description, supporting documents, expert names, and risk level labels.

[0013] Furthermore, the method constructs an SQL retrieval statement based on the entities extracted from the application materials to retrieve historical similar cases from the historical database.

[0014] Furthermore, the entities extracted from the application materials include the item name, composition description, and usage description; Before constructing the SQL retrieval statement based on the entities extracted from the application materials, the entities extracted from the application materials are also standardized. This standardization process includes synonym normalization, term abbreviation expansion, and multilingual unified conversion.

[0015] Furthermore, the similarity is calculated based on the cosine similarity between vectors.

[0016] Furthermore, the method also includes: Based on the risk feature nodes corresponding to the obtained application materials, query the risk level criterion node corresponding to the risk feature node. If there is an updated text description relative to the risk level criterion node in the corresponding enhanced context, it is determined to be new evidence for the risk feature node, and the information is extracted and added to the corresponding risk level criterion node.

[0017] Furthermore, the method also includes: Based on the risk feature node corresponding to the obtained application materials, query the risk level criterion node corresponding to the risk feature node. If there is an update record of the relevant content in the risk level criterion node in the corresponding enhanced context, then the corresponding content is added to the risk level criterion node and the corresponding most recent verification timestamp is updated.

[0018] Furthermore, the method obtains the differentiated risk status, confidence level, and risk level criteria of each component of the declared materials under different risk dimensions based on the acquired risk feature node set and its confidence level and risk level criterion nodes, so as to process the declared materials.

[0019] Compared with existing technologies, this invention organically integrates a special item classification system, a risk factor system, and risk criteria by constructing a hierarchical risk factor knowledge graph. Furthermore, based on dynamic context-enhanced entity linking technology, it achieves precise matching from application materials to knowledge graph nodes, offering the following advantages: (1) More refined risk feature expression: This invention extracts features-level entities and sets up a meaningful combination paradigm of special item subclass nodes × risk factor nodes to form risk feature nodes for risk level judgment; in the process of matching application materials with hierarchical risk factor knowledge graph, a hierarchical graph matching idea is proposed, and a three-granularity matching strategy is adopted: first, the special item subclass is matched with coarse granularity through Top-K screening; then, the risk factor is matched with medium granularity through threshold filtering; and finally, the risk feature nodes are matched with fine granularity. This realizes the matching search process of the above combination paradigm, which can finely depict the differentiated risk status of each component of composite special items under different risk dimensions. Compared with manual keyword retrieval, the evaluation granularity and accuracy are more in line with the actual review needs.

[0020] (2) Explicit and auditable risk knowledge: All risk judgment basis in this invention is stored in the form of structured nodes, which constitute risk level judgment nodes and correspond to risk feature nodes, including literature sources and expert signatures, etc. Regulatory agencies can review them one by one, overcoming the black box problem of large model fine-tuning schemes and meeting the compliance requirements of high responsibility scenarios in port supervision.

[0021] (3) Knowledge updates do not require retraining: The addition and modification of the graph nodes of the present invention take effect immediately, supporting rapid response to emerging infectious disease pathogens or new biological agents, and the update cost is far lower than that of large model retraining schemes.

[0022] (4) Eliminating the risk of illusion: The criteria of this invention are strictly derived from real documents and expert opinions. Entity links are based on vector similarity matching rather than text generation, so no risk evidence of non-existence will be generated.

[0023] (5) The present invention integrates current declaration information with historical cases to construct enhanced context, which can provide substantial semantic support for subsequent context enhancement and improve the accuracy of risk identification results for special items entering and leaving the port. Attached Figure Description

[0024] Figure 1 This is a flowchart illustrating a method for identifying risks of special items entering and leaving ports based on knowledge graph construction and entity linking, as provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of a special item risk knowledge graph structure provided in an embodiment of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0026] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0027] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0028] Example 1 like Figure 1 As shown, this embodiment provides a method for risk identification of special items entering and leaving ports based on knowledge graph construction and entity linking, including the following steps: S0: Based on the risk identification requirements of special items entering and leaving the port, a hierarchical risk factor knowledge graph is constructed in advance, and the corresponding content source text is obtained. The hierarchical risk factor knowledge graph has a multi-layer structure, including a root node for special items, a sub-category node for special items, a risk factor node, a risk feature node, and a risk level criterion node. Relevant information on special items and risk factors is extracted from the content source text and mapped to the nodes of each layer of the hierarchical risk factor knowledge graph. In this embodiment, as Figure 2 As shown, the hierarchical risk factor knowledge graph adopts a five-layer node structure, specifically including: The first layer is the root node for special items, representing the total category of all special items entering and leaving the country; The second layer consists of special item sub-category nodes, organized according to a parent-child structure based on the official classification directory. Each sub-category node contains a node identifier, classification definition, basic item characteristic description (including the names and attributes of each functional component), and inherent risk level attribute. The third layer consists of risk factor nodes, including but not limited to pathogenicity, toxicity, infectivity, drug resistance, environmental persistence, host range, and ecological security risks. The fourth layer consists of risk feature nodes, which are generated by a meaningful combination of special item sub-category nodes and risk factor nodes. The validity judgment is based on expert knowledge and historical assessment records. The fifth layer is the risk level judgment node, which includes a set of keywords, text descriptions, supporting documents, expert names, and high / medium / low risk level labels.

[0029] The knowledge graph's content originates from four types of data: official special item classification lists and regulatory documents, historical risk assessment records (expert comments and approval opinions), academic literature in the field of biosafety, and accumulated application materials from previous applications. Information extraction technology is used to identify item names, risk descriptions, evidence citations, and expert conclusions, mapping the extracted results to corresponding graph node attributes.

[0030] Information extraction from the source text includes: The source text is subjected to entity recognition, relation extraction, and evidence-conclusion pair extraction. Entity recognition is used to obtain item names, ingredient keywords, and risk terms. Relationship extraction is used to identify the association between items and risk factors and the basis for judging the effectiveness of the combination. Evidence-conclusion pairs are used to extract the keywords, risk level labels, literature sources, and expert signatures on which each risk judgment is based.

[0031] The information extracted from the source text is mapped to four layers of nodes in a hierarchical risk factor knowledge graph, including: Based on the entity recognition results, supplement the special item sub-category nodes with classification definitions, basic item feature descriptions, and inherent risk level attributes; supplement the risk factor nodes with the applicable scope of each risk factor (based on high-frequency words in historical assessments). Based on the relationship extraction results, the effectiveness of the combination between special item subclass nodes and risk factor nodes is confirmed, and risk feature nodes are generated. Based on the evidence-conclusion analysis, supplement the risk level criterion nodes with the required keyword set, text description, supporting documents, expert names, and risk level labels.

[0032] Optionally, the hierarchical risk factor knowledge graph is continuously and dynamically updated through three mechanisms: Case-driven updates: node replenishment is triggered after each new approval case is completed. Literature-driven updates: Regularly retrieve the latest biosafety literature and add new evidence to the corresponding criteria nodes; Expert-driven revision allows domain experts to directly revise risk level labels and criteria. All revision operations are recorded, including revision time, reviser, and revision basis, and version traceability is supported.

[0033] For novel items not included in the map in the application materials, candidate sub-category nodes are automatically generated and the expert confirmation process is triggered. After confirmation, the node is activated and the combination effectiveness judgment with each risk factor node is initiated.

[0034] S1: Obtain the application materials for the risks to be identified, extract entities from the application materials, retrieve historical similar cases based on the extracted entities, and construct an enhanced context that includes the application materials and their historical similar cases; vectorize the enhanced context as a whole; In this embodiment, SQL retrieval statements are dynamically generated based on entities extracted from the application materials, similar cases are retrieved from the historical database, and an enhanced context containing current application information and historical cases is constructed; the enhanced context is vectorized using a domain fine-tuning embedding model.

[0035] The specific process includes: S11: First, entity identification and standardization are performed: key entities such as product name, component description, and usage instructions are extracted from the application materials. Complex items are decomposed at the component level. Standardization processes include synonym normalization, expansion of terminology abbreviations, and unified conversion across multiple languages.

[0036] S12: Then, perform dynamic historical retrieval: Generate SQL queries based on standardized entities and retrieve similar cases from the historical approval database. Employ fuzzy matching and reverse time ordering strategies, prioritizing the return of recent cases. For complex items, each component is retrieved independently, and the results are merged to form a historical case pool.

[0037] S13: Context Enhancement and Vectorization: Enhanced context is constructed by integrating current application information with historical cases. A domain-fine-tuned embedding model is used to generate semantic vectors, achieving implicit encoding of historical knowledge.

[0038] S2: Based on the vectorized enhanced context, a three-granularity matching strategy is adopted: coarse-grained matching of specific item subclasses (Top-K filtering), medium-granularity matching of risk factors (threshold filtering), and fine-grained matching of risk feature nodes, outputting a set of risk feature nodes and the link confidence of each node. The link confidence is the cosine similarity between the enhanced context vector and the risk feature node vector.

[0039] The final output is a set of risk feature nodes and their confidence levels.

[0040] The specific process includes: matching the vectorized enhanced context with the information vectors of all special item subclass nodes respectively, obtaining one or more special item subclass nodes with the highest similarity, and adding the corresponding risk factor nodes to the candidate subclass set; matching the vectorized enhanced context with the information vectors of each risk factor node respectively within the candidate subclass set, and selecting the candidate risk factor set based on the similarity calculation results. After pairing the elements in the candidate subclass set and the candidate risk factor set one by one, multiple query combinations are formed. For each query combination, the existence of a corresponding risk feature node is queried. Based on the query results, the set of risk feature nodes and the information of their confidence and risk level criterion nodes are output as the risk identification results of special items entering and leaving the port.

[0041] Based on the acquired risk feature node set and its confidence level and risk level criterion node information, the differentiated risk status, confidence level and risk level criterion of each component of the declared materials under different risk dimensions are obtained, so as to process the declared materials, such as resisting the entry of special items with high risk level, and providing relevant risk level criterion.

[0042] Preferably, the method also includes a knowledge graph reverse supplementation process, which can be: Based on the risk feature nodes corresponding to the obtained application materials, query the risk level criterion nodes corresponding to the risk feature nodes. If there is an updated text description relative to the risk level criterion node in the corresponding enhanced context, it is determined as new evidence for the risk feature node, and the information is extracted and added to the corresponding risk level criterion node.

[0043] Based on the risk feature nodes corresponding to the obtained application materials, query the risk level criterion nodes corresponding to the risk feature nodes. If there are update records of relevant content in the risk level criterion nodes in the corresponding enhanced context, then supplement the corresponding content into the risk level criterion nodes and update the corresponding most recent verification timestamp.

[0044] Below is an example of entity links and reverse knowledge graph complementation for a certain recombinant lentiviral vector: background A company has applied to import a certain recombinant lentiviral vector for gene expression research. The application materials state that the product uses a four-plasmid packaging system, including the envelope plasmid VSV-G, the packaging plasmid gag-pol, the helper plasmid rev, and a transfer plasmid containing the target gene. The declared purpose is scientific research, and the source is a foreign research institution. The following describes the complete execution process of this example method.

[0045] 1: Entity Recognition and Standardization The system extracts the overall entity e0 (recombinant lentiviral vector, as a subclass matching anchor) and four component-level entities from the application materials: e1 (VSV-G, envelope plasmid), e2 (gag-pol, packaging plasmid), e3 (rev, helper plasmid), as well as the use entity e4 (gene expression research) and the normalized use category entity e5 (scientific research).

[0046] 2: Dynamic historical search Using each entity as a search term, the system generates separate SQL queries to retrieve similar cases from the historical approval database. Taking component entity e1 (VSV-G) as an example, the query conditions are that the item name or ingredient description contains the keyword "VSV-G," returning the ten most recent records sorted in reverse chronological order by approval date. Separate queries are generated for e2 (gag-pol) and e0 (recombinant lentivirus vector), each returning relevant historical cases. The merged search results form a historical case pool H, where each record includes the historical item name, expert evaluation opinions, and the final risk assessment result, providing substantial semantic support for subsequent contextual enhancement.

[0047] 3: Context Enhancement and Vectorization The current application information (item name, description of four plasmid components, and research purpose) is concatenated with the historical case pool H to construct the enhanced context C_aug. The historical case paragraphs contain several structured records with expert opinions. The domain-fine-tuned embedding model vectorizes C_aug into v_aug, which simultaneously encodes the semantic features of the currently applied item and historical evaluation experience, used for subsequent similarity calculations with knowledge graph node vectors. The enhanced contexts of each component entity are independently vectorized and then merged to ensure that the semantic features of each component are fully expressed.

[0048] 4: Hierarchical graph matching In the coarse-grained matching stage, the cosine similarity between v_aug and the node vectors of all special item subclasses is calculated. The recombinant virus vector node has the highest similarity, and its components are added to the candidate subclass set C.

[0049] In the medium-granularity matching stage, the similarity between v_aug and each risk factor node is calculated within the range C. The set F of candidate risk factors with a similarity of not less than the threshold θ1 includes: host range (semantically driven by VSV-G host tropism), replication ability (semantically driven by gag-pol packaging plasmids and four-plasmid systems), biosafety level (driven by overall item characteristics), and usage compliance (driven by usage entity). Infectivity and drug resistance are not included in F* because their semantic similarity with the current item is less than θ1.

[0050] In the fine-grained matching phase, the system queries the knowledge graph for each pair of combinations in C and F. The results are as follows: Combination (VSV-G, host range) has a corresponding risk feature node r1 in the graph, conf(r1)=0.87, and is included; Combination (VSV-G, replication capability) has a node r2, conf(r2)=0.82, and is included; Combination (gag-pol, replication capability) matches node r3, conf(r3)=0.79, and is included; Combination (recombinant viral vector, biosafety level) has a node r4, conf(r4)=0.75, and is included; Combination (recombinant viral vector, usage compliance) has a node r5, conf(r5)=0.71, and is included. For combinations (VSV-G, infectivity) and (gag-pol, drug resistance), there are no corresponding valid nodes in the graph, so they are automatically skipped, and no linking results are forcibly generated.

[0051] Where conf is the augmented context vector v_aug and the risk feature node embedding vector v_r i The cosine similarity is used as the link confidence of the node: conf(r i = cos(v_aug, v_r) i ); This corresponds to the description in step S2: "The link confidence is the cosine similarity between the enhanced context vector and the risk feature node vector."

[0052] The final output risk feature node set F = {r1, r2, r3, r4, r5}, link confidence Conf = {0.87, 0.82, 0.79, 0.75, 0.71}.

[0053] 5: Knowledge Graph Reverse Supplementation After the entity link is completed, the system checks the graph status of each risk feature node in the link result and identifies two types of situations that need to trigger reverse supplementation: one is the addition of new evidence to existing nodes, and the other is the corroborative update.

[0054] For node r1 (VSV-G × host range), the system queries the criterion node for this node and finds that its keyword set already includes terms such as "wide host tropism" and "infectious host range." However, a recent approval record was found in the historical case pool H, whose expert evaluation opinion provides an updated textual description of the host range risk of VSV-G and cites a relevant literature published in 2023. The system determines that this constitutes new evidence for this node and triggers a literature-driven update: the title, source, and expert opinion summary of the literature are added to the literature source field and expert opinion field of the corresponding criterion node r1. At the same time, the item name, approval result, and expert evaluation opinion of this application case are added to the historical approval database for subsequent retrieval.

[0055] For node r4 (recombinant viral vector × biosafety level), the system query revealed that the description of BSL-2 operation requirements in this node's criterion came from records several years ago. The current application materials clearly state the laboratory operation level requirements, which are consistent with current regulatory standards, thus serving as corroborating evidence of the existing criterion. The system added the corresponding evidence fragments from the current application materials to the keyword set of this node's criterion and updated the most recent verification timestamp of the criterion, reflecting the status of the criterion's content being reconfirmed by this case.

[0056] All the above reverse supplementation operations record the operation time, trigger source (case number in this application), and operation type (new evidence / evidence update / new feature), supporting version traceability. All revisions are marked as pending review by the system before being formally written into the graph, and are activated after confirmation by domain experts to ensure that the professional accuracy of the graph content is not reduced due to automated supplementation.

[0057] It should be stated that the entry and exit declaration materials involved in this plan are data to be processed obtained from legally public data interfaces or authorized data, and these data themselves have legal source attributes.

[0058] Furthermore, the risk identification method for special items entering and leaving ports described in this invention targets data packets that have met the data entry and exit security assessment conditions stipulated by relevant laws and regulations, or that have met the statutory cross-border compliance procedures. In other words, this risk identification method for special items entering and leaving ports applies to legally cross-border transferable data; this solution does not include the processing of non-compliant data.

[0059] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for risk identification of special goods entering and leaving ports based on knowledge graph construction and entity linking, characterized in that, Includes the following steps: Based on the risk identification requirements for special items entering and leaving the port, a hierarchical risk factor knowledge graph is constructed in advance, and the corresponding content source text is obtained. The hierarchical risk factor knowledge graph has a multi-layer structure, which includes a root node for special items, a sub-category node for special items, a risk factor node, a risk feature node, and a risk level criterion node. Relevant information on special items and risk factors is extracted from the content source text and mapped to the nodes of each layer of the hierarchical risk factor knowledge graph. Obtain the application materials for the risks to be identified, extract entities from the application materials, retrieve historical similar cases based on the extracted entities, construct an enhanced context containing the application materials and their historical similar cases, and perform overall vectorization processing; The vectorized enhanced context is matched with the information vectors of all special item subclass nodes respectively, and one or more special item subclass nodes with the highest similarity are obtained. The corresponding risk factor nodes are added to the candidate subclass set. Within the candidate subclass set, the vectorized enhanced context is matched with the information vectors of each risk factor node respectively. Based on the similarity calculation results, the candidate risk factor set is selected. After pairing the elements in the candidate subclass set and the candidate risk factor set one by one, multiple query combinations are formed. For each query combination, the existence of a corresponding risk feature node is queried. Based on the query results, the risk feature node set and its confidence level and risk level criterion node information are output as the risk identification result of special items entering and leaving the port.

2. The method for risk identification of special items entering and leaving ports based on knowledge graph construction and entity linking as described in claim 1, characterized in that, The root node for special items represents the general category of special items for entry and exit. The special item subclass node is a child node of the corresponding special item root node, and includes node identifier, classification definition, basic item feature description and inherent risk level attribute; The risk factor nodes include pathogenicity, toxicity, infectivity, drug resistance, environmental persistence, host range, and ecological safety risks. The risk feature node is generated by combining special item sub-category nodes and risk factor nodes based on expert knowledge and historical assessment records; The risk level criterion nodes include a set of keywords, textual descriptions, supporting documents, expert names, and risk level labels.

3. The method for risk identification of special items entering and leaving ports based on knowledge graph construction and entity linking as described in claim 2, characterized in that, Relevant information on special items and risk factors is extracted from the text source of the content, including: The source text of the content is subjected to entity recognition, relation extraction, and evidence-conclusion pair extraction. Entity recognition is used to obtain the item name, component keywords, and risk terms. Relationship extraction is used to identify the association between items and risk factors and the basis for judging the effectiveness of the combination. Evidence-conclusion pairs are used to extract the keywords, risk level labels, literature sources, and expert signatures on which each risk judgment is based.

4. The method for risk identification of special items entering and leaving ports based on knowledge graph construction and entity linking as described in claim 3, characterized in that, The information extracted from the source text is mapped to the nodes of each layer of the hierarchical risk factor knowledge graph, including: Based on the entity recognition results, supplement the special item subclass nodes with classification definitions, basic item feature descriptions, and inherent risk level attributes; supplement the risk factor nodes with the scope of application for each risk factor; Based on the relationship extraction results, the effectiveness of the combination between special item subclass nodes and risk factor nodes is confirmed, and risk feature nodes are generated. Based on the evidence-conclusion analysis, supplement the risk level criterion nodes with the required keyword set, text description, supporting documents, expert names, and risk level labels.

5. The method for risk identification of special items entering and leaving ports based on knowledge graph construction and entity linking as described in claim 1, characterized in that, The method constructs an SQL retrieval statement based on entities extracted from the application materials to retrieve historical similar cases from a historical database.

6. The method for risk identification of special items entering and leaving ports based on knowledge graph construction and entity linking as described in claim 5, characterized in that, The entities extracted from the application materials include the item name, composition description, and usage description; Before constructing the SQL retrieval statement based on the entities extracted from the application materials, the entities extracted from the application materials are also standardized. This standardization process includes synonym normalization, term abbreviation expansion, and multilingual unified conversion.

7. The method for risk identification of special items entering and leaving ports based on knowledge graph construction and entity linking as described in claim 1, characterized in that, The similarity is calculated based on the cosine similarity between vectors.

8. The method for risk identification of special items entering and leaving ports based on knowledge graph construction and entity linking as described in claim 1, characterized in that, The method further includes: Based on the risk feature nodes corresponding to the obtained application materials, query the risk level criterion node corresponding to the risk feature node. If there is an updated text description relative to the risk level criterion node in the corresponding enhanced context, it is determined to be new evidence for the risk feature node, and the information is extracted and added to the corresponding risk level criterion node.

9. The method for risk identification of special items entering and leaving ports based on knowledge graph construction and entity linking as described in claim 1, characterized in that, The method further includes: Based on the risk feature node corresponding to the obtained application materials, query the risk level criterion node corresponding to the risk feature node. If there is an update record of the relevant content in the risk level criterion node in the corresponding enhanced context, then the corresponding content is added to the risk level criterion node and the corresponding most recent verification timestamp is updated.

10. The method for risk identification of special items entering and leaving ports based on knowledge graph construction and entity linking as described in claim 1, characterized in that, The method obtains the differentiated risk status, confidence level, and risk level criteria of each component of the declared materials under different risk dimensions based on the acquired risk feature node set and its confidence level and risk level criterion nodes, so as to process the declared materials.

Images