A method and system for risk assessment of entry and exit articles at a port based on framework mining

Abstract

The application relates to a kind of port entry and exit article risk assessment method and system based on framework mining, method includes: obtaining the historical evaluation conclusion text of port entry and exit article and carrying out text preprocessing, obtain dimension label sequence;Dimension label sequence is carried out hierarchical clustering and corresponding dimension set is constructed, obtain multiple representative framework;Obtain port entry and exit article information, and select optimal evaluation framework;Dimension-by-dimension evidence matching and risk quantification are carried out;The process of evidence matching and risk quantification includes: if the evidence corresponding to dimension exists, then according to evidence polarity and historical risk determination reference, the risk score of dimension is calculated;If the evidence corresponding to dimension does not exist, then trigger penalty mechanism, improve the risk score of dimension;Based on dimension risk score, calculate comprehensive risk score, and carry out comprehensive evaluation to port entry and exit article information.Compared with prior art, the application has the advantage that the special article appearing newly can be distinguished.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and natural language processing technology, and in particular to a risk assessment method and system for inbound and outbound goods at ports based on frame mining. Background Technology

[0002] Risk assessment of special goods entering and leaving ports is a complex judgment task that highly relies on professional knowledge and historical experience. In the assessment process, experts not only need to identify the risk characteristics of the goods, but also need to examine multiple risk dimensions sequentially according to a certain logical framework, and form a comprehensive judgment based on the evidentiary status of each dimension. This assessment framework—that is, which risk dimensions should be examined for a specific type of goods, in what order, and what type of evidence is needed to support each dimension—is tacit knowledge accumulated by experts over a long period of time. Currently, it still exists in the form of unstructured text in historical assessment conclusions and has not yet been systematically extracted and reused. Existing technologies have the following main shortcomings: (1) The evaluation framework relies on manual definition and is difficult to cover all item types. Existing intelligent evaluation methods based on rule engines have their evaluation frameworks defined by human experts from top to bottom, which is costly to build and difficult to maintain.

[0003] (2) Cross-type experience cannot be effectively transferred. Current assessment methods establish an independent assessment system for each type of special item. The risk dimension knowledge shared between different item types cannot be systematically identified and reused, resulting in a waste of knowledge assets.

[0004] (3) Lack of assessment capabilities for new and special items. New items not covered by the existing assessment system are constantly appearing in the application materials. The assessment of new and special items relies on manual review by experts, which seriously restricts the improvement of customs clearance efficiency for such special items.

[0005] (4) The framework knowledge contained in historical assessment conclusions has not been fully utilized. A large number of completed risk assessment conclusion texts record the complete reasoning process of experts examining each risk dimension in sequence. The implicit assessment framework logic contained therein has not yet been systematically extracted and structured expressed by automated means.

[0006] The invention disclosed in CN121563447B is an intelligent review method and system for consumer rights protection based on a large language model, belonging to the fields of artificial intelligence and natural language processing. The method includes: preprocessing the text to be reviewed to obtain structured text feature data; using a pre-trained large language model to perform semantic understanding on the structured text feature data to obtain text semantic vectors containing domain semantic information; jointly retrieving and matching the text semantic vectors using a knowledge graph and a rule base to identify and output a set of risk elements; and making a compliance determination based on the set of risk elements to obtain a preliminary review conclusion including risk level, sufficiency of evidence, rule matching degree, and baseline confidence level. This invention automatically transforms the original text into a highly credible review conclusion with interpretability through intelligent processing throughout the entire process. However, the evaluation framework constructed by this scheme relies on manual definition, resulting in poor coverage and robustness for new types of issues. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the existing technology by providing a risk assessment method and system for inbound and outbound goods at ports based on frame mining. By performing structured analysis and pattern mining on historical assessment conclusion texts, the invention automatically extracts assessment frames for different special goods types from the bottom up, and transfers the assessment frames of known types to new or rare goods based on the similarity of goods features, thereby realizing structured risk assessment and interpretable report generation for any type of special goods.

[0008] The objective of this invention can be achieved through the following technical solutions: A risk assessment method for inbound and outbound goods at ports of entry based on frame mining includes: The historical assessment conclusion text of goods entering and leaving the port is obtained and preprocessed to obtain a dimensional label sequence; the comprehensive similarity between sequence pairs in the dimensional label sequence is calculated; based on the comprehensive similarity, the dimensional label sequence is hierarchically clustered and a corresponding dimensional set is constructed to obtain multiple representative frames; the representative frame is an assessment template composed of a determined dimensional set. Obtain information on inbound and outbound goods at the port; calculate and weight the similarity of biological attributes, functional uses, and existing evaluation dimensions between the information on inbound and outbound goods at the port and the information on goods in the representative framework, and obtain a comprehensive feature similarity; select the optimal evaluation framework based on the comprehensive feature similarity. Based on the aforementioned optimal evaluation framework, evidence matching and risk quantification are performed on the port entry and exit goods information dimension by dimension. The evidence matching and risk quantification process includes: if evidence corresponding to a dimension exists, then the corresponding dimension risk score is calculated based on the evidence polarity and historical risk judgment reference for each dimension; if evidence corresponding to a dimension does not exist, then a penalty mechanism is triggered to increase the dimension risk score based on the preset dimension risk score; based on the dimension risk score, a comprehensive risk score is calculated; and based on the comprehensive risk score, the port entry and exit goods information is comprehensively evaluated.

[0009] Furthermore, the risk scores of each dimension are normalized and weighted to obtain a preliminary risk score; based on the preliminary risk score, a constraint check is performed to obtain the comprehensive risk score.

[0010] Furthermore, the formula for calculating the normalized weighted sum is: S = Σ(w i ·s i ) / Σw i w i = f i · c i f i = n i / N Where S is the preliminary risk score, w i As the weight of the risk dimension, s i For the evidence polarity of the risk dimension, f i For the dimension frequency weights, c i n represents the proportion of the DFP pairing formed by the current item feature and this dimension relative to all DFP pairs for this item feature. i N represents the number of texts containing this dimension, and N is the total number of texts in the cluster.

[0011] Furthermore, the similarity of biological attributes, functional uses, and existing evaluation dimensions between the port entry / exit goods information and the goods information in the representative framework is calculated and weighted to obtain the comprehensive feature similarity. The corresponding calculation formula is as follows: Sim(A,B) = α·sim_bio + β·sim_use + γ·sim_dim Where Sim(A,B) represents the comprehensive feature similarity, α, β and γ are weight coefficients, sim_bio represents the biological attribute similarity, sim_use represents the functional use similarity, and sim_dim represents the existing evaluation dimension similarity.

[0012] Furthermore, the text preprocessing process includes: Identify the evaluation dimension category to which each sentence in the historical evaluation conclusion text belongs; extract keywords for each dimension based on the evaluation dimension category to obtain a standardized keyword set; Based on the standardized keyword set, evidence fragments for each dimension are classified by evidence polarity to obtain corresponding three-category polarity labels and confidence levels; based on the three-category polarity labels and confidence levels, the dimension label sequence is constructed; the three-category polarity labels include supporting low risk, supporting high risk, and evidence missing.

[0013] Furthermore, a two-layer dimension annotation is employed to identify the evaluation dimension category to which each sentence in the historical evaluation conclusion text belongs; the two-layer dimension annotation includes: Using a pre-built domain dictionary, the coarse-grained dimension affiliation of each sentence is determined; based on the coarse-grained dimension affiliation, a pre-built large language model is invoked to perform fine-grained dimension boundary identification to obtain the dimension boundary position, which is used to identify and evaluate the dimension category.

[0014] Furthermore, the comprehensive similarity is obtained by calculating and weighting the degree of overlap of dimension sets and normalized edit distance between sequence pairs in the dimension label sequence.

[0015] Furthermore, the dimension set includes mandatory dimensions and optional dimensions; the dimension is determined to be a mandatory dimension or an optional dimension based on the frequency of its occurrence within the corresponding representative framework.

[0016] Furthermore, based on the comprehensive feature similarity, the optimal evaluation framework is selected. The corresponding selection process includes: Using the comprehensive feature similarity as a weight, the weighted union of the dimension sets of multiple representative frames is taken to form an initial frame; based on the port entry and exit goods information, effective evaluation dimension-item feature combinations are identified, and evaluation dimensions that do not appear in the initial frame are added to the initial frame to obtain the optimal evaluation frame.

[0017] The present invention also provides a system for a risk assessment method for inbound and outbound goods at ports based on frame mining, comprising a memory and a processor, wherein the memory stores a computer program, and the processor calls the computer program to execute the steps of any of the methods described above.

[0018] Compared with the prior art, the present invention has the following advantages: (1) This invention calculates the dimension risk score based on the evidence polarity and historical risk judgment reference of each dimension. It also considers the universality of the dimension in the evaluation of similar items and the specificity of the specific item feature. This makes the dimensions that appear frequently in the evaluation of similar items and have strong specificity for the specific features of the item dominate in the comprehensive risk score. It effectively integrates the two orthogonal dimensions of universality and specificity, so that the final score result has higher risk discrimination and business semantic consistency. By calculating similarity for each type of special item from three different perspectives, cross-type framework retrieval not only has biological security guarantees, but also compliance alignment at the business scenario level, and empirical support at the historical data level. Ultimately, this translates into higher retrieval accuracy, stronger generalization ability for unknown items, and more granular risk differentiation capabilities in port assessment practices.

[0019] (2) This invention calculates similarity for each type of special item from three different perspectives. Biological attribute similarity ensures a deep characterization of the essential features related to biosafety, functional use similarity ensures alignment with business application scenarios, and existing assessment dimension similarity provides empirical similarity references using historical experience. The weighted combination of the three makes cross-type framework retrieval no longer rely on simple matching of a single dimension, but integrates information from three levels: intrinsic attributes, extrinsic uses, and historical experience, which significantly improves the accuracy and generalization ability of the retrieval results, and ultimately can retrieve the most suitable assessment framework for each item.

[0020] (3) When calculating the similarity between two dimensional label sequences, this invention does not use a single metric, but comprehensively considers the complementary characteristics of the two dimensions. Jaccard similarity is used to evaluate the degree of overlap between the two sequences at the dimensional set level, that is, to focus on which dimensions are present; normalized edit distance is used to measure the consistency of the two sequences in the dimensional arrangement order, that is, to focus on the order in which the dimensions appear. Compared to a single similarity measure that uses only set overlap or only sequence order, this weighted combination can more comprehensively and delicately depict the true similarity of two assessment texts in terms of risk consideration logic. It will not misjudge texts with the same dimension set as dissimilar due to different order, nor will it ignore the different assessment logic reflected by the order difference due to high set overlap, thus significantly improving the semantic rationality and business interpretability of hierarchical clustering.

[0021] (4) When calculating the dimensional weights of the comprehensive risk score, the present invention adopts a two-factor product design; the first factor is the dimensional frequency weight, which reflects the degree of universality of the dimension in the evaluation of similar items, and the second factor is the historical co-occurrence importance of DFP, which reflects the degree of specificity of the DFP pair for the specific characteristics of the item. This makes the dimensions that appear frequently in the evaluation of similar items and have strong specificity for the specific characteristics of the item dominate in the comprehensive risk score, effectively integrating the two orthogonal dimensions of universality and specificity, so that the final score result has higher risk discrimination and business semantic consistency. Attached Figure Description

[0022] Figure 1 This is a flowchart of a risk assessment method for inbound and outbound goods at ports based on frame mining, provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the overall process of a risk assessment method for inbound and outbound goods at ports based on frame mining provided in an embodiment of the present invention. It shows the complete processing chain starting from historical assessment conclusion text, through text preprocessing, frame mining, similarity calculation and frame adaptation, risk quantification and constraint checking, and report generation, as well as the feedback loop of continuously incorporating new assessment conclusions into historical corpus to trigger frame updates. Figure 3 This is a schematic diagram of the assessment framework mining process for a risk assessment method for inbound and outbound goods at ports based on framework mining, provided in an embodiment of the present invention. Figure 4 This is a schematic diagram of frame migration based on the similarity of goods features, which is a risk assessment method for inbound and outbound goods at ports based on frame mining provided in an embodiment of the present invention. Figure 5 This is a schematic diagram of the dimension-by-dimensional risk quantification and constraint inspection process of a risk assessment method for inbound and outbound goods at ports based on frame mining, provided in an embodiment of the present invention. Figure 6 This is a schematic diagram of a structured assessment report format for a risk assessment method for inbound and outbound goods at ports based on frame mining, provided in an embodiment of the present invention. Detailed Implementation

[0023] 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.

[0024] 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.

[0025] 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.

[0026] Definitions: Jaccard similarity is a statistical metric used to measure the degree of similarity between two sets. Its value is equal to the size of the intersection of the two sets divided by the size of the union of the two sets. The value of this metric ranges from 0 to 1: when the two sets are completely identical, the ratio is 1; when the two sets have no common elements, the ratio is 0.

[0027] Normalized edit distance: This is a standardized form of edit distance used to measure the similarity between two strings. Edit distance refers to the minimum number of single-character editing operations required to transform one string into another; these operations typically include insertion, deletion, and replacement.

[0028] DFP matching refers to the process of pairing and identifying the characteristics of an item to be evaluated with valid evaluation dimensions confirmed in historical evaluations. Specifically, the system analyzes the subcategories or component characteristics of the item to be evaluated, such as human cell components, and then retrieves risk dimensions that are validly associated with that characteristic from the historical evaluation knowledge base, such as cytotoxicity, forming a meaningful combination unit, namely "item characteristic × evaluation dimension", i.e., DFP.

[0029] It should be stated that the port entry and exit goods data 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.

[0030] Furthermore, the cross-border data transfer method described in this application targets data packets that have met the data export security assessment conditions stipulated by relevant laws and regulations or have completed the statutory cross-border transfer compliance procedures. In other words, this transfer method is applied to legally transferable cross-border data, and this solution does not include any steps for circumventing regulatory oversight in the transfer of non-compliant data.

[0031] Example 1 like Figure 1 As shown, this embodiment provides a risk assessment method for inbound and outbound goods at ports based on frame mining. The method includes the following steps: S1: Obtain the historical assessment conclusion text of goods entering and leaving the port and perform text preprocessing to obtain the dimension label sequence; calculate the comprehensive similarity between sequence pairs in the dimension label sequence; based on the comprehensive similarity, perform hierarchical clustering on the dimension label sequence and construct the corresponding dimension set to obtain multiple representative frames; the representative frame is an assessment template composed of the determined dimension set; S101: Preprocessing of Historical Assessment Conclusions Text: The historical assessment conclusion text is segmented and its dimensional attribution is identified to form a structured sequence representation with dimensional labels and evidence polarity. This specifically includes three processing steps: (1) Evaluation Dimension Identification: A two-layer annotation mechanism combining coarse-grained domain dictionary attribution and fine-grained boundary identification of inter-sentence dependencies is adopted to identify the evaluation dimension category to which each sentence belongs. The evaluation dimension categories include: inherent biological risk of the item, pathogen detection and inactivation status, country of origin and supplier qualification, transport packaging compliance, declaration of use constraints, and evidence integrity.

[0032] (2) Keyword extraction and normalization: Keywords in each dimension are extracted using a domain dictionary combined with TF-IDF statistics. Synonyms, abbreviations and multilingual expressions are disambiguated and normalized to form a standardized set of keywords.

[0033] (3) Evidence Polarity Identification: For each dimension of evidence fragments, a large language model classification method based on cue word engineering is used to identify the evidence polarity, including: supporting low risk / supporting high risk / evidence missing. Specifically, the evidence fragments of each dimension, together with their respective dimension labels and domain background cue, are constructed as structured input. The domain fine-tuning language model is called to output the three-class polarity labels and confidence scores. For fragments with confidence scores below the threshold, the polarity word matching rules based on the domain dictionary are used for supplementary judgment. Finally, a dimension-level evidence-conclusion pair sequence is formed, which serves as the input for the framework mining in step two.

[0034] S102: Evaluation Framework Mining like Figure 3 As shown, this paper demonstrates three sub-steps for extracting dimensional sequences from historical evaluation texts: sequence clustering and typical framework induction, as well as methods for identifying required and optional dimensions. Using the dimensional label sequence of each evaluation text as input, typical evaluation frameworks for each item type are extracted from the bottom up through similarity calculation, hierarchical clustering, and representative framework induction.

[0035] Sequence similarity calculation: First, dimensional sequence extraction is performed, such as extracting evaluation text t1, evaluation text t2, and evaluation text t3, etc. Each evaluation text includes multiple different dimensions, such as dimension D1, dimension D2, dimension D3, dimension D4, and dimension D5. Then, the similarity between any two dimensional sequences is calculated by weighted combination of comprehensive Jaccard similarity, which represents the degree of overlap of the dimensional sets, and normalized edit distance, which represents the consistency of the dimensional order.

[0036] Hierarchical clustering: Based on the sequence similarity matrix, historical evaluation texts are hierarchically clustered. Each cluster corresponds to a category of item evaluation scenarios. For example, cluster C1 corresponds to the evaluation of the same type of items, and cluster C2 corresponds to the evaluation of another type of items.

[0037] Representative framework summary: Extract a representative evaluation framework for each cluster—dimensions with a frequency of occurrence not lower than the first threshold θ are marked as mandatory dimensions; dimensions with a frequency between the second threshold θ2 and the first threshold θ are marked as optional dimensions; dimensions with a frequency lower than the second threshold θ2 are not included in the framework.

[0038] S2: Obtain information on inbound and outbound goods at the port; calculate and weight the similarity of biological attributes, functional uses, and existing evaluation dimensions between the information on inbound and outbound goods at the port and the information on goods in the representative framework, and obtain a comprehensive feature similarity; select the optimal evaluation framework based on the comprehensive feature similarity. S201: Establish feature similarity between item types: For each type of special item, three feature vectors are constructed for cross-type frame retrieval: (1) Biological attribute vector: Encodes the biological origin, type of active ingredient, genetic material characteristics, etc. of the item, using biosafety ontology hierarchical coding; (2) Functional purpose vector: Encodes typical declared purpose categories, mapped according to the port declaration purpose standard classification system; (3) Existing evaluation dimension set vector: Encodes the set of dimensions involved in historical evaluations, such as a binary sparse vector.

[0039] The comprehensive feature similarity is a weighted combination of the cosine similarity of the three types of vectors: Sim(A,B) = α·sim_bio +β·sim_use + γ·sim_dim, where α, β, and γ are adjustable weight coefficients, and the default values ​​are set based on the verification results of the historical framework migration effect.

[0040] S202: Evaluation framework adaptation for the item to be evaluated: like Figure 4As shown, this demonstrates the three-dimensional feature vectors of the item to be evaluated and known item types, namely biological attributes, functional uses, and existing evaluation dimension sets; the comparison and comprehensive similarity calculation; and the complete process of frame adaptation, such as dimension supplementation and pruning. The metadata output of the adapted frame is also annotated. Specific steps include: Based on feature similarity, the top K most similar known item types are retrieved, and frame merging, dimension augmentation, and dimension pruning are performed sequentially to generate a customized evaluation framework for the item to be evaluated. (1) Frame merging: Take the weighted union of the dimension sets of K frames to form the initial frame. The weights are determined by the similarity between the corresponding item type and the item to be evaluated.

[0041] (2) Dimension Supplementation: The evaluation dimensions of the item to be evaluated are combined with the item characteristics. That is, the unique risk dimensions identified in the DFP matching that are not yet covered by the initial framework are added to the framework and labeled as "item-specific dimensions". The DFP consists of meaningful pairings between the item's sub-category or component characteristics and risk dimensions. Its validity is determined based on the frequency of occurrence of the pairing in historical evaluations and the results of expert confirmation.

[0042] (3) Dimension trimming: Remove dimensions that are deemed to have no substantial meaning to the items to be evaluated in the DFP matching from the frame to avoid irrelevant dimensions affecting the evaluation conclusion.

[0043] (4) Degradation: For new items whose similarity to all known types is below the lower limit threshold θ4, degrade them to a general basic framework covering all required dimensions, mark them as conservative estimates, and recommend initiating an expert manual review process.

[0044] S3: Based on the optimal evaluation framework, conduct dimensional evidence matching and risk quantification of information on inbound and outbound goods at ports of entry. The process of evidence matching and risk quantification includes: if evidence for a dimension exists, calculate the corresponding dimension risk score based on the evidence polarity and historical risk judgment reference for each dimension; if evidence for a dimension does not exist, trigger a penalty mechanism to increase the dimension risk score based on the preset dimension risk score; calculate the comprehensive risk score based on the dimension risk score; and conduct a comprehensive evaluation of information on inbound and outbound goods at ports of entry based on the comprehensive risk score.

[0045] S301: Dimensional Risk Quantification like Figure 5 As shown, the execution order of the four stages—evidence matching, penalty for missing evidence (i.e., differentiated handling of necessary and auxiliary evidence), weighted summation, and constraint checks—is illustrated, highlighting the key design principle that constraint checks take precedence over the weighted summation result. Evidence matching and risk quantification are performed dimension-by-dimensionally based on the adaptation framework, with constraint checks applied. Specific steps include: (1) Evidence matching and scoring: When evidence exists, risk scores are calculated based on evidence polarity and historical risk determination reference dimensions; (2) Penalty for missing evidence: A penalty mechanism is triggered when evidence is missing—when necessary evidence is missing, the risk score is adjusted to the 90th percentile of the historical distribution, and when auxiliary evidence is missing, it is adjusted to the 75th percentile. The distinction between necessary and auxiliary evidence is determined based on the frequency weight of this dimension in the framework.

[0046] (3) Weighted summation: Normalized weighted summation of risk scores for each dimension S = Σ(w i ·s i ) / Σw i , where w i = f i · c i Dimensional frequency weight f i f represents the frequency of this dimension within its respective cluster. i = n i / N, where the numerator is the number of texts containing that dimension, and the denominator is the total number of texts in the cluster; DFP historical co-occurrence importance c i c represents the proportion of the DFP pairing formed by the current item feature and this dimension relative to all DFP pairs for this item feature. i = m i / M i Where the numerator is the historical occurrence count of the specific DFP, and the denominator is the sum of the occurrence counts of all DFPs involving the item feature. Since f i This reflects the universality of dimensions in the evaluation of similar items, c i This reflects the specificity of the DFP pair for the specific characteristics of the item, and the product of the two ensures that the high-frequency and highly specific dimensions dominate in the weighting.

[0047] (4) Constraint check: For a predefined set of high-risk triggering conditions, including uninactivated highly pathogenic pathogens, positive detection of key pathogens, and involvement in the list of controlled biological agents, if any triggering condition exists in the application materials, the comprehensive risk level will be forcibly upgraded to high risk. The execution priority of this constraint is higher than the weighted summation result.

[0048] S302: Generation of Structured Assessment Report: like Figure 6 As shown, the natural language version of the report, including evaluation paragraphs for each dimension, evidence anchoring, evidence missing information, comprehensive conclusions, and evaluation meta-information, is presented alongside the machine-readable JSON version. The parallel structure of the two versions, along with the field correspondence between them, is demonstrated. The parallel dual-version of the natural language report and the machine-readable JSON report is generated according to the evaluation framework dimensions. Specific steps include: (1) Natural Language Report: Each dimension corresponds to an assessment paragraph, which includes an assessment conclusion sentence, evidence summary, DFP identifier and evidence polarity description; the evidence missing dimension clearly marks the missing type and supplementary verification prompts; the report ends with a comprehensive risk level statement, disposal recommendations and assessment meta-information; the assessment meta-information includes framework source, adaptation operation record and constraint check results.

[0049] (2) JSON report: Each field is bound to the evaluation dimension label and DFP identifier, supporting automatic archiving and statistical analysis; the field structure corresponds one-to-one with each paragraph of the natural language report.

[0050] (3) Continuous optimization of the framework library: Each completed evaluation conclusion can trigger the re-mining of the framework after being included in the historical corpus; the similarity calculation weight coefficients α, β, and γ can be adaptively adjusted according to the feedback of the framework migration effect. Specifically, by comparing the migration framework coverage rate, i.e. the proportion of dimensions in the adapted framework that are confirmed as effective by experts, and the dimension accuracy rate, i.e. the proportion of unpruned dimensions to the total number of dimensions in the initial framework, the weighted average of the two indicators on the validation set is used as the optimization target. The grid search is used to update the values ​​of α, β, and γ in the preset discrete value range. The adjustment results are recorded in the framework library meta-information and take effect in the next batch, forming a closed-loop mechanism for continuous optimization.

[0051] Example 2 This embodiment provides a novel framework for evaluating gene-edited cell products, encompassing a complete evaluation and migration process, including: A company has applied to import a novel gene-edited cell product, named GECPXXXXB. This product type has no directly corresponding historical evaluation record in the existing framework library. The product is prepared from human cells transduced with a lentiviral vector and then genetically modified, containing residual viral vector components. The declared intended use is for scientific research and preclinical studies, and the source is an overseas research institution. The following is... Figure 2 The flowchart illustrates the complete execution process of the method of the present invention.

[0052] Step 1: Preprocessing of Historical Evaluation Conclusions Text The system retrieves historical evaluation conclusion text samples related to the product's application materials from the historical evaluation database. Each text undergoes a two-layer dimensional annotation: first, a domain dictionary is used to assign coarse-grained dimensions to each sentence; the domain dictionary contains biological terminology, regulatory expressions, and risk description phrases; then, a large language model is used to identify fine-grained dimensional boundaries for each sentence: consecutive sentence groups within a sliding window, along with the coarse-grained assignment results, are constructed as prompt words. The model determines whether adjacent sentences have undergone dimensional switching, outputting the dimensional boundary position and corresponding dimensional label, thus completing the fine-grained boundary identification.

[0053] Taking a historical assessment conclusion of a recombinant viral vector as an example, the following dimensional label sequence is output after double-layer annotation: [component pathogenicity, replication capacity, insertion gene safety, biosafety level, use claim constraints, transport and packaging compliance]. Simultaneously, the evidence polarity of each dimension is identified: the component pathogenicity dimension polarity is "supports low risk," such as replication-defective vectors; the replication capacity dimension polarity is "supports low risk," such as split packaging; and the use claim constraints dimension polarity is "lack of evidence," such as no purchase contract provided. Synonyms are normalized; for example, replication-defective and self-inactivating are uniformly mapped to the standard terminology of replication defects, forming a standardized dimension-polarity sequence, which serves as input for subsequent framework mining.

[0054] Step 2: Evaluation Framework Mining Dimensional label sequences of 128 historical assessment conclusion texts were extracted from the historical database. The comprehensive similarity between all sequence pairs was calculated, with a Jaccard similarity weight of 0.6 and a normalized edit distance weight of 0.4. Agglomerated hierarchical clustering was performed on the similarity matrix using the Ward link criterion, truncating the cluster tree at height h=0.35, resulting in 7 clusters. The two larger clusters correspond to "Recombinant Viral Vectors" (containing 38 texts) and "Blood and Blood Products" (containing 26 texts), respectively.

[0055] A representative framework was extracted from the clustering of "recombinant viral vectors": the frequency of each dimension within the cluster is as follows: pathogenicity of components (0.97), replication capacity (0.92), biosafety level (0.89), usage declaration constraints (0.84), safety of inserted genes (0.76), transport and packaging compliance (0.71), epidemic situation in the country of origin (0.39), and host range (0.31). Using θ=0.70 as the mandatory threshold and θ2=0.30 as the optional threshold, the first six dimensions constitute the mandatory dimension set, while the epidemic situation in the country of origin and host range constitute the optional dimension set.

[0056] Step 3: Feature Similarity Calculation Three types of feature vectors are constructed for GECPXXXXB: biological attribute vector encoding "human cells / gene modification / containing viral vector components"; functional use vector encoding "scientific research / preclinical research"; and existing evaluation dimension set vector initially empty, such as novel items.

[0057] The overall similarity was calculated with known item types in the framework library, where α=0.5, β=0.3, γ=0.2: the biological attribute similarity sim_bio=0.79 with the recombinant viral vector, indicating a high match in viral vector components; the functional use similarity sim_use=0.88; and the overall similarity Sim=0.83, the highest similarity. The overall similarity with blood products Sim=0.61; and the overall similarity with microorganisms Sim=0.47. The recombinant viral vector framework with the highest similarity was selected as the initial framework, and the blood product framework was included as a second reference, where K=2.

[0058] Step 4: Evaluate Framework Adaptability Frame merging: The weighted union of K=2 frames is taken with similarity as the weight to form an initial frame containing 6 mandatory dimensions and 2 optional dimensions; the recombinant viral vector frame is the main one, and the detection status dimension of blood products is included in the optional set with a lower weight.

[0059] Dimension Supplementation: Perform DFP matching on GECPXXXXB to identify "human cell components × cytotoxicity" as a valid evaluation dimension-item feature combination, i.e., DFP. The corresponding cytotoxicity evaluation dimension was not covered in the initial framework and was supplemented and labeled as "item-specific dimension".

[0060] Dimensional clipping: DFP matching results show that the combination "GECPXXXXB×drug resistance" has never appeared in historical assessments. Experts determined that this combination has no substantial significance for this type of item, and the corresponding drug resistance assessment dimension has been removed from the framework.

[0061] The adapted framework includes seven dimensions: pathogenicity of components (mandatory), replication capacity (mandatory), biosafety level (mandatory), usage declaration constraints (mandatory), safety of inserted genes (mandatory), transport and packaging compliance (mandatory), and cytotoxicity (item-specific dimension, mandatory); country of origin epidemic situation (optional), and blood product testing status (optional, from the second reference framework). Framework metadata records: migrated from the recombinant viral vector framework (Sim=0.83), with the cytotoxicity dimension added and the drug resistance dimension removed.

[0062] Step 5: Dimensional Risk Quantification Based on the adaptation framework, evidence matching and risk quantification were performed on each dimension of the application materials: (1) Pathogenicity of components: The application materials state that the carrier is a replication defect type, the evidence polarity is low risk, and the dimension risk score = 0.2; (2) Replication capability dimension: Three-plasmid packaging system, replication defect, evidence polarity is low risk, score = 0.15; (3) Insertion gene safety dimension: The application materials state that the target gene is a fluorescent reporter gene, with no known toxicity and low risk, score = 0.15; (4) Cytotoxicity dimension (item-specific): The application materials did not provide a cytotoxicity test report, triggering the penalty for missing necessary evidence, and the risk score was raised to the 90th percentile of the historical distribution = 0.68; (5) Declaration of Purpose Constraint Dimension: The application materials did not provide a purchase contract and cooperation agreement, triggering the penalty for missing necessary evidence, and the risk score was raised to the 90th percentile = 0.65; (6) Transport packaging compliance dimension: Complies with UN3373 standards, low risk, score = 0.10; (7) Biosafety level dimension: BSL-2 operation requirements are clearly stated in the materials, low risk, score = 0.18.

[0063] Weighted summation: The weights of each dimension are determined by multiplying the frequency weight by the importance of the DFP (Discretionary Factor in Risk Assessment). The initial comprehensive risk score after weighting is 0.41. Constraint checks: There are no mandatory triggering conditions such as uninactivated highly pathogenic pathogens or controlled biological agents, and no mandatory increase in risk level. The comprehensive risk score is mapped to the medium risk level.

[0064] Step Six: Generation of Structured Assessment Report The system generates a structured evaluation report according to the dimensional order of the adaptation framework, such as... Figure 5 As shown: The natural language version generates an assessment paragraph for each dimension. Missing evidence for the cytotoxicity and use declaration constraints dimensions is clearly marked, and a verification prompt is given suggesting that the applicant supplement the cytotoxicity test report and purchase contract. The overall risk level is medium risk, and the recommended course of action is "to approve entry, with the condition that the applicant supplement the cytotoxicity test report and contract documents."

[0065] The JSON version binds each field to the dimension label and DFP identifier. The report metadata is fully recorded: the framework has been migrated from recombinant viral vectors, a cytotoxicity dimension has been added, the drug resistance dimension has been pruned, and constraint checks have confirmed that no mandatory high-risk conditions have been triggered.

[0066] After review and confirmation, the assessment results were incorporated into the historical corpus, triggering a framework library update. The validity record of gene-edited cell product DFPs was updated after this case verification, making them available for direct use in framework adaptation for subsequent similar products.

[0067] The key technical points of this invention include: 1. A bottom-up approach to extracting evaluation frameworks from historical evaluation conclusion texts, including a dimension identification mechanism based on domain dictionary and inter-sentence dependency relationship dual-layer annotation, sequence similarity calculation (i.e., a weighted combination of Jaccard similarity and normalized edit distance), hierarchical clustering, and a frequency threshold identification mechanism for mandatory / optional dimensions.

[0068] 2. Evaluation Dimension - Item Feature Combination (DFP): Definition and validity determination method: The meaningful pairing of item sub-category or component features with risk dimensions is used as the basic unit. The validity of the combination is determined based on the frequency of occurrence in historical evaluations and expert confirmation, which serves as the decision basis for dimension addition and trimming in framework adaptation.

[0069] 3. A method for calculating the comprehensive feature similarity between item types based on the weighted combination of three-dimensional feature vectors, and a retrieval mechanism for a cross-type evaluation framework based on similarity.

[0070] 4. Evaluation framework adaptation methods, including weighted union of multiple frameworks, dimension addition and pruning based on DFP matching results, and conservative estimation processing mechanism for degenerating new items into a general basic framework.

[0071] 5. A penalty mechanism for missing evidence, including methods for distinguishing between necessary and supplementary evidence and differentiated risk score adjustment strategies, such as adjusting to the 90th / 75th percentile.

[0072] 6. Constraint and inspection mechanisms, including the definition of high-risk trigger condition sets and the guarantee mechanism that their execution priority is higher than the weighted summation result.

[0073] 7. A parallel dual-version generation method for structured evaluation reports, including natural language versions organized in order of framework dimensions, each judgment anchored to a DFP identifier, a complete record of evaluation meta-information including framework source, adaptation operation, constraint check status, and output format bound to a machine-readable JSON version.

[0074] 8. The framework library has a continuous optimization mechanism, including incorporating new evaluation conclusions into historical corpora to trigger framework re-mining, and an adaptive adjustment method for similarity calculation weight coefficients based on feedback from migration effects.

[0075] 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 risk assessment method for inbound and outbound goods at ports based on frame mining, characterized in that, include: The historical assessment conclusion text of goods entering and leaving the port is obtained and preprocessed to obtain a dimension label sequence; the comprehensive similarity between sequence pairs in the dimension label sequence is calculated. Based on the comprehensive similarity, hierarchical clustering is performed on the dimension label sequence to construct a corresponding dimension set, resulting in multiple representative frames; the representative frame is an evaluation template composed of a determined dimension set. Obtain information on inbound and outbound goods at the port; calculate and weight the similarity of biological attributes, functional uses, and existing evaluation dimensions between the information on inbound and outbound goods at the port and the information on goods in the representative framework, and obtain a comprehensive feature similarity; Based on the comprehensive feature similarity, the optimal evaluation framework is selected; Based on the optimal evaluation framework, evidence matching and risk quantification are performed on the information of inbound and outbound goods at the port of entry and exit in a dimensional manner. The process of evidence matching and risk quantification includes: if evidence corresponding to a dimension exists, then the corresponding dimension risk score is calculated based on the evidence polarity and historical risk judgment reference for each dimension; if evidence corresponding to a dimension does not exist, then a penalty mechanism is triggered to increase the dimension risk score based on the preset dimension risk score; a comprehensive risk score is calculated based on the dimension risk score; and a comprehensive assessment of the port entry and exit goods information is conducted based on the comprehensive risk score.

2. The method for risk assessment of inbound and outbound goods at ports based on frame mining as described in claim 1, characterized in that, The risk scores for each dimension are normalized and weighted to obtain a preliminary risk score; based on the preliminary risk score, a constraint check is performed to obtain the comprehensive risk score.

3. The method for risk assessment of inbound and outbound goods at ports based on frame mining as described in claim 2, characterized in that, The formula for calculating the normalized weighted summation is as follows: S = Σ(w i ·s i ) / Σw i w i = f i · c i f i = n i / N c i = m i / M Where S is the preliminary risk score, w i As the weight of the risk dimension, s i For the evidence polarity of the risk dimension, f i For dimensional frequency weights, DFP is the pairing of item features and evaluation dimensions, c i n represents the proportion of the DFP formed by the current item feature and this dimension relative to all DFPs for that item feature. i The number of texts containing this dimension is N, where N is the total number of texts in the cluster, and m is the number of texts in the cluster. i M represents the historical occurrence count of the DFP formed by the current item feature and this dimension, and M is the sum of the occurrence counts of all DFPs in which this item feature participates.

4. The method for risk assessment of inbound and outbound goods at ports based on frame mining as described in claim 1, characterized in that, The similarity of biological attributes, functional uses, and existing assessment dimensions between the port entry / exit goods information and the goods information in the representative framework is calculated and weighted to obtain the comprehensive feature similarity. The corresponding calculation formula is as follows: Sim(A,B) = α·sim_bio + β·sim_use + γ·sim_dim Where Sim(A,B) represents the comprehensive feature similarity, α, β and γ are weight coefficients, sim_bio represents the biological attribute similarity, sim_use represents the functional use similarity, and sim_dim represents the existing evaluation dimension similarity.

5. The method for risk assessment of inbound and outbound goods at ports based on frame mining as described in claim 1, characterized in that, The text preprocessing process includes: Identify the evaluation dimension category to which each sentence in the historical evaluation conclusion text belongs; extract keywords for each dimension based on the evaluation dimension category to obtain a standardized keyword set; Based on the standardized keyword set, evidence fragments for each dimension are classified by evidence polarity to obtain corresponding three-category polarity labels and confidence levels; based on the three-category polarity labels and confidence levels, the dimension label sequence is constructed; the three-category polarity labels include supporting low risk, supporting high risk, and evidence missing.

6. The method for risk assessment of inbound and outbound goods at ports based on frame mining as described in claim 5, characterized in that, A two-layer dimension annotation is used to identify the evaluation dimension category to which each sentence in the historical evaluation conclusion text belongs; the two-layer dimension annotation includes: Using a pre-built domain dictionary, the coarse-grained dimension affiliation of each sentence is determined; based on the coarse-grained dimension affiliation, a pre-built large language model is invoked to perform fine-grained dimension boundary identification to obtain the dimension boundary position, which is used to identify and evaluate the dimension category.

7. The method for risk assessment of inbound and outbound goods at ports based on frame mining as described in claim 1, characterized in that, The comprehensive similarity is obtained by calculating and weighting the overlap of dimension sets and normalized edit distance between sequence pairs in the dimension label sequence.

8. The method for risk assessment of inbound and outbound goods at ports based on frame mining as described in claim 1, characterized in that, The set of dimensions includes mandatory dimensions and optional dimensions; a dimension is determined to be a mandatory dimension or an optional dimension based on its frequency of occurrence within the corresponding representative framework.

9. The method for risk assessment of inbound and outbound goods at ports based on frame mining according to claim 1, characterized in that, Based on the comprehensive feature similarity, the optimal evaluation framework is selected. The selection process includes: Using the comprehensive feature similarity as a weight, the weighted union of the dimension sets of multiple representative frames is taken to form an initial frame; based on the port entry and exit goods information, effective evaluation dimension-item feature combinations are identified, and evaluation dimensions that do not appear in the initial frame are added to the initial frame to obtain the optimal evaluation frame.

10. A system for risk assessment of inbound and outbound goods at ports of entry based on frame mining, characterized in that, It includes a memory and a processor, the memory storing a computer program, the processor invoking the computer program to perform the steps of the method as described in any one of claims 1 to 9.

Images