A compliance procurement qualification auditing method based on semantic parsing and logical tree construction and electronic equipment
Patent Information
- Application Number
- CN202611301834.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-26
- Publication Date
- 2026-09-25
AI Technical Summary
[0003]本申请的主要目的在于提供一种基于语义解析与逻辑树构建的合规采购资质审核方法及电子设备，以解决现有文本比对系统无法理解证照文本中的排他逻辑语义以及对非一致性表述容错率低的缺陷
[0016]本申请提供的技术方案带来了以下有益效果：通过将非结构化的证照经营范围文本以及其中的限制性自然语言词汇解析为层级分明的合规逻辑树，准确且自动化地将包含、排除、仅限等人类复杂排他逻辑重构为计算机可执行的布尔运算表达式，规避了传统字面关键词匹配带来的违规误报漏洞；同时，通过建立多证照图谱语义等价性融合机制，利用融合字面差异与高维语义夹角的混合计算模型对跨文本实体进行等价性判定，有效消除了由行政区划更迭或企业缩写引起的底层数据冲突，增强了系统对地址、名称描述变化的容错率，实现了无人工干预下的精准审核控制，降低了医疗器械采购业务中的人工复核压力与合规风险。
Smart Images
Figure CN122820239A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of natural language processing technology, particularly to the use of computers to process natural language, specifically a compliance procurement qualification review method and electronic device based on semantic parsing and logic tree construction applied in a multi-certificate comparison scenario for medical devices. This technical solution achieves high-precision compliance verification in a multi-text interaction environment through deep application dependency parsing, logical modification entity joint extraction, semantic graph fusion, compliance logic tree construction, sentence vector similarity comparison, and semantic equivalence calculation. Background Technology
[0002] In medical device procurement, qualification verification requires checking various certificates and licenses provided by suppliers to confirm their consistency and strict matching with the items in the purchase order. Due to the complexity of medical device classification standards, the business scope descriptions on certificates often contain numerous restrictive or exclusive natural language expressions. Current mainstream automated verification systems use keyword matching algorithms based on literal strings for comparison, lacking an understanding of syntactic structure and unable to handle exclusive or restrictive logic, easily leading to false positives. Furthermore, conventional systems have extremely poor tolerance for cross-certificate entity descriptions; literal comparison algorithms cannot recognize the semantic equivalence of underlying relationships, easily causing comparison blockages. This results in a high system error rate, increases the workload of manual review, and makes it difficult to meet the needs of automated workflow. Summary of the Invention
[0003] The main purpose of this application is to provide a compliance procurement qualification review method and electronic device based on semantic parsing and logic tree construction, so as to solve the shortcomings of existing text comparison systems that cannot understand the exclusive logical semantics in the certificate text and have low tolerance for inconsistent expressions.
[0004] To achieve the above objectives, this application provides a compliance procurement qualification review method based on semantic parsing and logic tree construction, including: acquiring and preprocessing multi-source certificate texts and order texts to be reviewed; inputting the preprocessed multi-source certificate texts into a joint entity extraction network to extract business entities and logical modifiers, and outputting business entity labels and logical modifier label sequences; performing dependency parsing on text fragments containing the logical modifier label sequences, setting logical modifier nodes as parent nodes according to word dependency relationships, setting the business entities governed by the logical modifier nodes as child nodes, and constructing a compliance logic tree with Boolean logic relationships. The process involves: converting the scope of the certificates into Boolean expressions; extracting entity nodes from different multi-source certificate texts; calculating the semantic similarity between the entity nodes using a hybrid scoring model that integrates word vector cosine similarity and normalized edit distance; establishing corresponding semantic connectivity edges between the entity nodes in the multi-certificate semantic graph when the semantic similarity is greater than an equivalence threshold; extracting the parameters to be reviewed from the order text; traversing the multi-certificate semantic graph to locate the target compliance logic tree; substituting the parameters to be reviewed into the Boolean expressions of the target compliance logic tree for evaluation and reasoning; and outputting a judgment review report based on the evaluation results.
[0005] Furthermore, the process of acquiring and preprocessing multi-source certificate texts and order texts to be reviewed includes word segmentation and vectorization. The process of inputting the preprocessed multi-source certificate texts into a joint entity extraction network to extract business entities and logical modifiers, and outputting business entity labels and logical modifier label sequences, includes using a pre-trained language model to map the vectorized characters into word embedding vector sequences. The word embedding vector sequences are then input into a sequence labeling layer to identify medical device entities and natural language words representing restrictive or exclusive relationships. A preset logical labeling system is used to synchronously label the natural language words, including inclusion labels representing inclusion relationships, exclusion labels representing exclusive relationships, and restriction labels representing limiting relationships, generating the corresponding logical modifier label sequences.
[0006] Furthermore, the process of inputting the word embedding vector sequence into the sequence labeling layer to identify medical device entities and natural language words representing restriction or exclusivity includes inputting the word embedding vector sequence into a bidirectional long short-term memory network layer to extract forward and backward contextual feature vectors; concatenating the forward and backward contextual feature vectors and inputting them into a conditional random field layer to output a globally optimal entity and logical joint label sequence, so as to locate the boundary of the medical device entity and lock the natural language words that appear immediately adjacent to the medical device entity.
[0007] Furthermore, the process of constructing a compliance logic tree with Boolean logic relationships by setting logical modifier nodes as parent nodes and the business entities governed by the logical modifier nodes as child nodes based on word dependency relationships includes: extracting the dependency syntax tree from the sentence based on a dependency parser; traversing the dependency syntax tree and extracting the central words marked as logical modifiers as control nodes; searching downwards along the dependency arc and taking entity words that have a modification or domination relationship with the central words as controlled nodes; assigning Boolean logic operators to each controlled node according to the label attributes of the control nodes; when the label attribute is an exclusion label, assigning a NOT operator to the corresponding controlled node; when the label attribute is an inclusion label, assigning an AND operator to the corresponding controlled node, thereby generating the hierarchical compliance logic tree.
[0008] Furthermore, the process of converting the scope of certifications into a Boolean expression includes traversing the nodes of the compliance logic tree from bottom to top; converting the business entities corresponding to the leaf nodes into logical variables; converting the Boolean logical operators corresponding to the non-leaf nodes into logical operators; and concatenating the logical variables and the logical operators according to the hierarchical structure of the compliance logic tree to generate a computer-executable Boolean expression. The Boolean expression is used to output a false value signal for variables that do not meet the preset qualification range during logical operations.
[0009] Furthermore, the process of calculating the semantic similarity between entity nodes using a hybrid scoring model that integrates word vector cosine similarity and normalized edit distance includes obtaining the first word vector corresponding to the first certificate text entity node and the second word vector corresponding to the second certificate text entity node; calculating the cosine similarity between the first word vector and the second word vector; calculating the normalized edit distance between the literal strings corresponding to the first certificate text entity node and the second certificate text entity node; and linearly weighting the cosine similarity and the normalized edit distance according to a preset weight coefficient to output the semantic similarity.
[0010] Furthermore, the linear weighted summation calculation process is performed as follows: determining the similarity score.
[0011] ;
[0012] Where S is the semantic similarity; The preset weight coefficients are defined, with values ranging from 0 to 1; v1 and v2 represent the first word vector and the second word vector, respectively; CosSim(v1,v2) represents the cosine similarity of the angle between word vectors in the high-dimensional semantic space; s1 and s2 represent the literal strings of the first and second certificate text entity nodes, respectively; NED(s1,s2) represents the normalized edit distance characterizing the literal morphological differences. Used to convert differences into literal similarity scores.
[0013] Furthermore, the process of substituting the parameters to be reviewed into the Boolean operation expression of the target compliance logic tree for evaluation and reasoning, and outputting a judgment audit report based on the evaluation result, includes passing the parameters to be reviewed as input variables into the Boolean operation expression for node comparison; when the comparison hits a node with a non-operator, blocking the downward traversal of the logic tree and outputting a false value result representing a violation; in response to the false value result, generating the judgment audit report with a warning label, and triggering an interception action to block the flow of the order to be reviewed.
[0014] This application also provides an electronic device, including a memory and a processor, wherein the memory is configured to store a computer program, and the processor is communicatively connected to the memory and configured to execute the computer program to implement the aforementioned compliance procurement qualification review method based on semantic parsing and logic tree construction.
[0015] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the aforementioned compliance procurement qualification review method based on semantic parsing and logic tree construction.
[0016] The technical solution provided in this application brings the following beneficial effects: By parsing the unstructured business scope text of licenses and the restrictive natural language words therein into a hierarchical compliance logic tree, it accurately and automatically reconstructs complex human exclusion logic such as inclusion, exclusion, and limitation into computer-executable Boolean expressions, avoiding the loopholes of false positives caused by traditional literal keyword matching; at the same time, by establishing a semantic equivalence fusion mechanism for multi-license graphs, it uses a hybrid calculation model that integrates literal differences and high-dimensional semantic angles to determine the equivalence of cross-text entities, effectively eliminating underlying data conflicts caused by changes in administrative divisions or company abbreviations, enhancing the system's fault tolerance for changes in address and name descriptions, achieving precise review and control without human intervention, and reducing the pressure of manual review and compliance risks in medical device procurement. Attached Figure Description
[0017] Figure 1This is a flowchart of a compliance procurement qualification review method based on semantic parsing and logic tree construction provided in an embodiment of the present invention.
[0018] Figure 2 This is a system architecture diagram of the electronic device provided in the embodiments of the present invention.
[0019] Explanation of reference numerals in the attached figures
[0020] In the diagram: 201 - Processor, 202 - Memory, 203 - Communication Interface, V 201 - Electronic devices, V 202 -Bus architecture, V 203 - Front-end procurement system. Detailed Implementation
[0021] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.
[0022] In one embodiment of this application, a compliance procurement qualification review method based on semantic parsing and logic tree construction is provided. The underlying hardware architecture of this method is deployed on a server providing computing power and data storage support, and interfaces with the front-end interaction interface of a medical device procurement business system via a communication network to receive multi-source certificate images and structured procurement order data to be reviewed uploaded by procurement personnel. Through interactive coupling between a deep learning model and a logic engine running internally on the server, the conversion of natural language to Boolean expressions and automatic flow interception are achieved. This compliance procurement qualification review method includes the following implementation steps.
[0023] The server executes step S101, retrieving and preprocessing multi-source certificate texts and pending order texts. The system receives multi-source certificate texts after optical character recognition (OCR) and structured segmentation. These multi-source certificates include text information from business licenses, medical device operating licenses, medical device production licenses, and product registration certificates. It also receives the pending purchase order description text from the front-end business system. The preprocessing module performs word segmentation, part-of-speech tagging, and stop word removal on the aforementioned unstructured long text segments. To facilitate subsequent neural network computation, the preprocessing stage divides the text into regular sequences, obtaining the input certificate scope description text segment sequence. Among them, x n This represents the nth valid character or vocabulary unit after word segmentation and cleaning. By performing word segmentation and vectorization, redundant characters without practical business meaning in the text are removed, reducing the computational load on the subsequent neural network during the feature extraction stage.
[0024] In step S102, where the multi-source certificate text is input into the network and the output label sequence is generated, the server inputs the preprocessed multi-source certificate text into the joint entity extraction network to extract business entities and logical modifiers, and outputs a sequence of business entity labels and logical modifier labels. In this embodiment, the joint entity extraction network adopts a cascaded deep learning model architecture, and its execution logic includes the following stages: The vectorized characters are mapped to a word embedding vector sequence using a pre-trained language model. The above text segment sequence X is input into a BERT model fine-tuned with corpus data from the medical device domain to obtain a high-dimensional continuous word vector representation sequence. Among them, e n The word x represents n The corresponding high-dimensional word embedding vectors are then used. The word embedding vector sequence E is input into the sequence labeling layer to identify medical device entities and natural language words representing restrictive or exclusive characteristics. To accommodate long-distance text dependencies, the system inputs the word embedding vector sequence E into a bidirectional long short-term memory network layer for forward and backward contextual feature vector extraction. This bidirectional mechanism ensures that the network can capture both the restrictive words preceding a medical device word and its subsequent modifiers. After extracting the feature vectors, the forward and backward contextual feature vectors are concatenated and input into a conditional random field layer. The conditional random field constrains the global transition probability of the sequence labels, outputting a globally optimal entity and logical joint label sequence to locate the boundaries of medical device entities and identify natural language words immediately adjacent to them.
[0025] In this step, the system abandons the conventional approach of simply extracting noun entities and instead employs a pre-defined logical tagging system to synchronously annotate the aforementioned natural language vocabulary. The defined extraction tag set not only includes conventional system-defined tags, such as [ORG_NAME] for identifying company names, [LOC] for identifying addresses, and [PROD_CLASS] for identifying medical device category numbers, but also specifically adds logical modifier tags, including the inclusion tag [LOGIC_INCLUDE] representing inclusion relationships, the exclusion tag [LOGIC_EXCLUDE] representing exclusion relationships, and the restriction tag [LOGIC_LIMIT] representing restriction relationships. Through this synchronous annotation, not only is the location of business entities captured, but the location of negation and restriction semantic elements is also completed, providing structured anchor data for the subsequent generation of executable computation nodes.
[0026] In step S103, dependency parsing is performed to construct a compliance logic tree. Dependency parsing is conducted on text fragments containing the logical modifier tag sequence. Logical modifier nodes are set as parent nodes based on word dependency relationships, and the business entities governed by these logical modifier nodes are set as child nodes. A compliance logic tree with Boolean logic relationships is constructed to convert the scope of licenses into Boolean expressions. For text fragments with logical modifier tags extracted in the previous steps, the server calls the dependency parser to analyze the sentence structure and extract the dependency parsing tree from the sentence. The system traverses the dependency parsing tree and extracts the central words marked as logical modifiers as control nodes. Using these nodes as the core, the system searches downwards along the dependency arc, identifying entity words that have a modification or domination relationship with the central word as controlled nodes. Corresponding Boolean logic operators are assigned to each controlled node based on the tag attributes of the control nodes. When the label attribute is an exclusion label [LOGIC_EXCLUDE], the NOT operator is assigned to the corresponding controlled node; when the label attribute is an inclusion label [LOGIC_INCLUDE], the AND operator is assigned to the corresponding controlled node, thereby generating a hierarchical compliance logic tree.
[0027] To make the logic tree computable, the system defines a logic tree node Node=. Here, Type represents the operation or entity type of the node, Value represents the specific value, and Children represents the set of its child nodes. In the stage of converting the scope of licenses into Boolean expressions, the system traverses the nodes of the compliance logic tree from bottom to top. The business entities corresponding to leaf nodes are converted into logical variables, and the Boolean logical operators corresponding to non-leaf nodes are converted into logical operators. The logical variables and logical operators are then concatenated according to the hierarchical structure of the compliance logic tree to generate a computer-executable Boolean expression. For the text "Business Scope: Class II Medical Devices (excluding 6846)", the system constructs a two-level tree structure: the root node is inclusive, child node 1 is [PROD_CLASS]: Class II, and child node 2 is the exclusion parent node, and the child node of this exclusion node is [PROD_CLASS]: 6846. Following the aforementioned bottom-up concatenation logic, the system converts this natural language text into an executable Boolean expression:
[0028]
[0029] This Boolean expression is used to force a false value signal for variables that do not meet the preset qualification range during subsequent logical operations. Without the extraction and transformation of restrictive natural language words into computational nodes with Boolean attributes through the aforementioned sequence labeling and dependency parsing, the system would struggle to generate expression structures with executable properties, thus eliminating the crude drawbacks of literal matching.
[0030] The system calculates entity similarity and establishes semantic connectivity edges (S104). Entity nodes are extracted from different multi-source certificate texts. A hybrid scoring model fusing word vector cosine similarity and normalized edit distance is used to calculate the semantic similarity between entity nodes. When this semantic similarity is greater than an equivalence threshold, corresponding semantic connectivity edges are established between the entity nodes in the multi-certificate semantic graph. In the comparison of multiple certificates, the descriptions of addresses and company names often exhibit reasonable heterogeneity. The system constructs a multi-certificate semantic graph using the business license as the master node. When the address node extracted from the business license is compared with the address node of the common attribute to be compared in the business license, and it is found that the literal strings of the two are not completely equal, the semantic fusion calculation mechanism is initiated.
[0031] The system inputs two text nodes into the BERT model to obtain the first word vector v1 corresponding to the first certificate text entity node and the second word vector v2 corresponding to the second certificate text entity node. The cosine similarity between the first and second word vectors is calculated to measure semantic equivalence in the high-dimensional semantic space. Simultaneously, the normalized edit distance between the literal string s1 corresponding to the first certificate text entity node and the literal string s2 corresponding to the second certificate text entity node is calculated to characterize the literal morphological differences. The cosine similarity and the normalized edit distance are then linearly weighted and summed according to preset weight coefficients. The specific mathematical expression of this feature fusion-based semantic similarity hybrid scoring formula is as follows:
[0032]
[0033] Where S represents the semantic similarity output by the hybrid scoring model; This represents a preset weighting coefficient, with a value ranging from 0 to 1. In this embodiment, to balance semantic representation and literal form, the parameter weighting coefficient... The preferred value is 0.6. v1 and v2 represent the first and second word vectors, respectively; the first part of the formula, CosSim(v1,v2), calculates the cosine similarity of the angle between the word vectors in the high-dimensional semantic space. s1 and s2 represent the literal strings of the first and second certificate text entity nodes, respectively; NED(s1,s2) represents the normalized edit distance for calculating the literal morphological difference between the two strings, which measures the ratio of the minimum number of single-character edit operations required to convert one string into another to the maximum length of both; the normalized edit distance calculation in the second part is used to characterize the literal morphological difference, and subtracting 1 is to convert the difference into a literal similarity score with the same polarity as the cosine similarity.
[0034] The system will compare the calculated semantic similarity S with a preset semantic equivalence threshold. A comparison is made. In this embodiment, the semantic equivalence threshold is used. The preferred value is 0.85. When the semantic similarity is calculated using the hybrid scoring formula... When two addresses are deemed semantically equivalent, the system connects the business license node to the business license entity node via edges in the local graph database. By executing the above semantic fusion mechanism, underlying data conflicts across texts are eliminated, ensuring that the logic tree matching engine is not blocked by superficial literal differences and guaranteeing smooth association of underlying data.
[0035] The system proceeds to step S105, where it performs evaluation and reasoning using the substitution expression. It extracts the parameters to be reviewed from the order text, traverses the multi-certificate semantic graph to locate the target compliance logic tree, substitutes the parameters into the Boolean expression of the target compliance logic tree for evaluation and reasoning, and outputs a judgment and review report based on the evaluation result. The system extracts the product name and category code from the purchase order as parameters to be reviewed, and uses them as input variables in the Boolean expression for node comparison and evaluation. Since equivalent connected edges have been established at the underlying level, the system can accurately locate the compliance logic tree of the corresponding supplier. The system substitutes the order category code into this tree structure. If it falls into a branch of a NOT node, it indicates that the purchase item belongs to the excluded qualification range. At this time, the system blocks the downward traversal of the logic tree and outputs a false value representing a violation. In response to this false value, the system generates a judgment and review report with a clear reason for failure and a warning indicator, and triggers an interception action to block the flow of the order to be reviewed. For the common medical device statement "Agreed to operate Class II medical devices, but excluding implantable devices", when the purchase order query contains the parameter "orthopedic implant nail", Boolean operation judges it as a false value and automatically blocks it, effectively avoiding the structural loophole of traditional systems that mistakenly allow the order to proceed due to the capture of the keyword "implantable device".
[0036] Building upon the above embodiments, to ensure the stability and business collaboration efficiency of the compliance procurement qualification review system, in another embodiment of this application, the method further includes common-sense interface interaction and storage scheduling logic to support the overall business flow. After the server completes the construction of the local semantic graph, the system persistently stores the data through the mounted local graph database. When an interception action occurs, the server provides a front-end interaction and application programming interface communication protocol for the procurement business flow approval action, pushes the interception report to the manual review pool in real time, and freezes the corresponding funds for the current order. By adding a robust communication interface mechanism and storage mounting method, a high degree of collaboration between the compliance inference engine and the front-end business is ensured, achieving a closed loop in the business flow.
[0037] In one embodiment of this application, an electronic device is provided for implementing the above-described compliant procurement qualification review method. This electronic device mainly consists of a hardware computing unit and a storage component, and is deployed in an enterprise-level server system. The electronic device includes a memory 202 and a processor 201 communicatively connected to the memory 202. In practical applications, the electronic device also includes a communication interface 203 responsible for interacting with the front-end procurement system.
[0038] Memory 202 is configured to store the data source required for computer programs and the graph storage database. Memory 202 may be a high-speed random access memory, or may include non-volatile memory, such as at least one disk storage device, flash memory device, or other non-volatile solid-state storage device. Processor 201 serves as the control hub of the electronic device and is connected to other components via a bus structure. Processor 201 may be a general-purpose central processing unit, a microprocessor, an application-specific integrated circuit, or one or more graphics processing units and tensor processing units configured to execute relevant deep learning algorithms.
[0039] The processor 201 is configured to execute the computer program stored in the memory 202 to implement the following working logic: receiving multi-source certificate texts and purchase orders after optical character recognition via the communication interface 203; invoking computing power to perform preprocessing and joint entity extraction networks to extract business entities and logical modifiers; using a dependency parser to parse the syntax and construct a compliance logic tree, converting natural language expressions into Boolean expressions; simultaneously, the processor calculates the hybrid scoring similarity across certificate entities, fusing cosine similarity and normalized edit distance to establish a multi-certificate semantic graph; the processor 201 substitutes order parameters into the logical expressions for evaluation and reasoning, and when a negative node is encountered, it sends an interception signal containing the specific non-compliance location and a judgment review report via the communication interface 203. Through the coordinated operation of the various physical components of the above electronic device, a solid computing power and storage physical foundation is provided for the complex semantic parsing network and high-frequency Boolean evaluation operations.
[0040] In one embodiment of this application, a computer-readable storage medium is also provided, on which a computer program is stored. This computer-readable storage medium can be a solid-state drive or read-only memory embedded within a server, or it can be a removable storage medium. When the computer program is executed by a processor, it can fully implement the compliance procurement qualification review method based on semantic parsing and logic tree construction described in the foregoing embodiments. Its specific operating steps and logical reasoning process correspond completely to those of the foregoing method embodiments. By embedding an instruction set containing deep learning model reasoning logic and graph construction algorithms in the storage medium, complex natural language processing technologies can achieve standardized deployment and cross-platform migration.
[0041] This application's embodiments deeply integrate sequence labeling models with dependency parsing to parse and reconstruct unstructured business scope text of licenses and permits into a computable compliance logic tree. This allows the review system to overcome the bottleneck of literal matching and achieve automated verification and precise interception of complex human exclusive logical semantics. Simultaneously, relying on the semantic equivalence graph fusion calculation based on a hybrid model of word vector cosine similarity and normalized edit distance, the system enhances its fault tolerance and algorithm robustness when faced with inconsistent expressions of addresses and company names. These features work together to solve the pain point of heavy reliance on manual review in the comparison of multiple licenses and permits for medical devices, ensuring the rigor and compliance of the medical procurement process.
[0042] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.
Claims
1. A compliance procurement qualification review method based on semantic parsing and logic tree construction, characterized in that, This includes acquiring and preprocessing multi-source certificate texts and pending order texts; inputting the preprocessed multi-source certificate texts into a joint entity extraction network to extract business entities and logical modifiers, and outputting business entity labels and logical modifier label sequences; performing dependency parsing on text fragments containing the logical modifier label sequences, setting logical modifier nodes as parent nodes based on word dependency relationships, and setting the business entities governed by the logical modifier nodes as child nodes, constructing a compliance logic tree with Boolean logic relationships to transform the certificate scope into Boolean operation expressions; and extracting... For entity nodes in different multi-source certificate texts, semantic similarity between entity nodes is calculated using a hybrid scoring model that integrates word vector cosine similarity and normalized edit distance. When the semantic similarity is greater than an equivalence threshold, a corresponding semantic connection edge is established between the entity nodes in the multi-certificate semantic graph. The parameters to be reviewed are extracted from the order text to be reviewed, and the multi-certificate semantic graph is traversed to locate the target compliance logic tree. The parameters to be reviewed are substituted into the Boolean operation expression of the target compliance logic tree for evaluation and reasoning. Based on the evaluation result, a judgment review report is output.
2. The compliance procurement qualification review method based on semantic parsing and logic tree construction as described in claim 1, characterized in that, The process of acquiring and preprocessing multi-source certificate texts and order texts to be reviewed includes word segmentation and vectorization. The process of inputting the preprocessed multi-source certificate texts into a joint entity extraction network to extract business entities and logical modifiers, and outputting business entity labels and logical modifier label sequences, includes using a pre-trained language model to map the vectorized characters into word embedding vector sequences. The word embedding vector sequences are then input into a sequence labeling layer to identify medical device entities as business entities and natural language words representing restrictive or exclusive relationships. A preset logical labeling system is used to synchronously label the natural language words, including inclusion labels representing inclusion relationships, exclusion labels representing exclusive relationships, and restriction labels representing limiting relationships, generating the corresponding logical modifier label sequences.
3. The compliance procurement qualification review method based on semantic parsing and logic tree construction as described in claim 2, characterized in that, The process of inputting the word embedding vector sequence into the sequence labeling layer to identify medical device entities and natural language words representing restriction or exclusivity includes inputting the word embedding vector sequence into a bidirectional long short-term memory network layer to extract forward and backward contextual feature vectors; concatenating the forward and backward contextual feature vectors and inputting them into a conditional random field layer to output a globally optimal entity and logical joint label sequence, so as to locate the boundary of the medical device entity and lock the natural language words that appear immediately adjacent to the medical device entity.
4. The compliance procurement qualification review method based on semantic parsing and logic tree construction as described in claim 1, characterized in that, The process of setting logical modifier nodes as parent nodes and the business entities governed by the logical modifier nodes as child nodes, and constructing a compliance logic tree with Boolean logic relationships, includes extracting the dependency syntax tree from the sentence based on a dependency syntax parser. Traverse the dependency syntax tree and extract the central words marked as logical modifiers as control nodes; search downwards along the dependency arc and select entity words that have a modifying or dominating relationship with the central words as controlled nodes; assign Boolean logical operators to each controlled node according to the label attributes of the control nodes. When the label attribute is an exclusion label, assign the NOT operator to the corresponding controlled node; when the label attribute is an inclusion label, assign the AND operator to the corresponding controlled node, thereby generating the hierarchical compliance logic tree.
5. The compliance procurement qualification review method based on semantic parsing and logic tree construction as described in claim 4, characterized in that, The process of converting the scope of licenses into a Boolean expression includes traversing the nodes of the compliance logic tree from bottom to top; converting the business entities corresponding to the leaf nodes into logical variables; converting the Boolean logical operators corresponding to the non-leaf nodes into logical operators; and concatenating the logical variables and logical operators according to the hierarchical structure of the compliance logic tree to generate a computer-executable Boolean expression. The Boolean expression is used to output a false value signal for variables that do not meet the preset qualification range during logical operations.
6. The compliance procurement qualification review method based on semantic parsing and logic tree construction as described in claim 1, characterized in that, The process of calculating the semantic similarity between entity nodes by using a hybrid scoring model that integrates word vector cosine similarity and normalized edit distance includes obtaining the first word vector corresponding to the first certificate text entity node and the second word vector corresponding to the second certificate text entity node; and calculating the cosine similarity between the first word vector and the second word vector. Calculate the normalized edit distance between the literal strings corresponding to the first certificate text entity node and the second certificate text entity node; linearly weight the cosine similarity and the normalized edit distance according to a preset weight coefficient, and output the semantic similarity.
7. The compliance procurement qualification review method based on semantic parsing and logic tree construction as described in claim 6, characterized in that, The linear weighted summation calculation process is performed in the following manner to determine the similarity score. ； Where S is the semantic similarity; The preset weight coefficients are defined, with values ranging from 0 to 1; v1 and v2 represent the first word vector and the second word vector, respectively; CosSim(v1,v2) represents the cosine similarity of the angle between word vectors in the high-dimensional semantic space; s1 and s2 represent the literal strings of the first and second certificate text entity nodes, respectively; NED(s1,s2) represents the normalized edit distance characterizing the literal morphological differences. Used to convert differences into literal similarity scores.
8. The compliance procurement qualification review method based on semantic parsing and logic tree construction as described in claim 1, characterized in that, The process of substituting the parameters to be reviewed into the Boolean operation expression of the target compliance logic tree for evaluation and reasoning, and outputting a judgment audit report based on the evaluation result, includes: passing the parameters to be reviewed as input variables into the Boolean operation expression for node comparison; when the comparison hits a node with a non-operator, blocking the downward traversal of the logic tree and outputting a false value result representing a violation; in response to the false value result, generating the judgment audit report with a warning label, and triggering an interception action to block the flow of the order to be reviewed.
9. An electronic device, characterized in that, The system includes a memory and a processor, the memory being configured to store a computer program, and the processor being communicatively connected to the memory and configured to execute the computer program to implement the compliance procurement qualification review method based on semantic parsing and logic tree construction as described in any one of claims 1 to 8.
10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the compliance procurement qualification review method based on semantic parsing and logic tree construction as described in any one of claims 1 to 8.