Scientific and technological achievement intelligent matching method and system based on multi-dimensional semantic analysis and knowledge graph

Abstract

The invention relates to the technical field of intelligent matching, and discloses a scientific and technological achievement intelligent matching method and system based on multi-dimensional semantic analysis and a knowledge graph. The method comprises the following steps: extracting a technical parameter dictionary, a demand keyword and a demand industry label in a demand text, and constructing demand structured data; performing word segmentation and coding on the demand structured data to obtain a demand semantic vector, constructing an achievement semantic vector of each achievement in an achievement database, and calculating a cosine distance, an industry association distance and an IPC classification distance; the cosine distance, the industry association distance and the IPC classification distance are subjected to weighted calculation, a comprehensive matching score corresponding to each achievement is obtained, a target achievement list is screened based on the comprehensive matching scores, and a recommendation result is returned. And an accurate and efficient intelligent matching service is provided for transformation of scientific and technological achievements.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent matching, in particular to a scientific and technological achievement intelligent matching method and system based on multi-dimensional semantic analysis and knowledge graph. BACKGROUND

[0002] The transformation of scientific and technological achievements is a key link to promote the transformation of innovation achievements into real productivity, and the precise docking of demanders and suppliers directly affects the transformation efficiency. The traditional scientific and technological achievement matching platform adopts keyword search method, which can only match the literal same technical expressions, for example: when the demander describes "high-strength lightweight material" and the supplier describes "carbon fiber reinforced polymer", the matching fails due to the difference in vocabulary, thereby causing the problem of mismatching of the recommended results, affecting the success rate of transformation. SUMMARY

[0003] The present application provides a scientific and technological achievement intelligent matching method and system based on multi-dimensional semantic analysis and knowledge graph, which ensures the technical feasibility, industry adaptability and patent matching degree of the recommended results, and provides precise and efficient intelligent matching service for the transformation of scientific and technological achievements.

[0004] In a first aspect, the present application provides a scientific and technological achievement intelligent matching method based on multi-dimensional semantic analysis and knowledge graph, which comprises: extracting technical parameter dictionary, demand keywords and demand industry label in demand text, and constructing demand structured data; performing word segmentation and coding on the demand structured data to obtain demand semantic vector, and constructing achievement semantic vector of each achievement in achievement database; calculating the cosine distance of the demand semantic vector and the achievement semantic vector, calculating the industry correlation distance of the demand industry label and the achievement industry label, and mapping the demand keywords to IPC classification and achievement IPC classification to calculate IPC classification distance; weighting the cosine distance, the industry correlation distance and the IPC classification distance to obtain the comprehensive matching score corresponding to each achievement, and screening the target achievement list based on the comprehensive matching score and returning the recommended results.

[0005] In combination with the first aspect, in a first implementation manner of the first aspect of the present application, the extraction of technical parameter dictionary, demand keywords and demand industry label in demand text, and the construction of demand structured data, comprises: extracting the numerical value and unit of technical parameters from the demand text to obtain the technical parameter dictionary; performing keyword recognition on the demand text to obtain demand keywords, and mapping the demand keywords to demand industry label according to a preset industry classification mapping table; The requirement text, the technical parameter dictionary and the requirement industry label are packaged into a request body structure, sent to a BERT encoding interface through an HTTP POST method carrying an API authentication Token, and requirement structured data is obtained.

[0006] In combination with the first aspect, in a second implementation manner of the first aspect of the application, the requirement structured data is segmented and encoded to obtain a requirement semantic vector, and a result semantic vector of each result in the result database is constructed, including: A tokenizer segmenter of a BERT-base model is called to perform WordPiece segmentation on the requirement text in the requirement structured data, cut into a token sequence and add a CLS mark and a SEP mark at the beginning and end of the sequence to obtain a requirement token sequence; Each token in the requirement token sequence is mapped into a token matrix through token embedding, segment embedding and position embedding of an embedding layer; Multi-head self-attention calculation is performed on the token matrix to obtain a target output matrix, and a requirement semantic vector corresponding to the CLS mark is extracted from the target output matrix; A BERT encoding interface is called to generate a result semantic vector according to each result in the result database.

[0007] In combination with the first aspect, in a third implementation manner of the first aspect of the application, the cosine distance between the requirement semantic vector and the result semantic vector is calculated, the industry correlation distance between the requirement industry label and the result industry label is calculated, and the IPC classification distance between the requirement IPC classification and the result IPC classification is calculated, including: The cosine distance between the requirement semantic vector and the result semantic vector is calculated; The shortest relationship path is queried according to the requirement industry label and the result industry label, the path hop count is obtained and divided by a preset maximum path length to obtain the industry correlation distance; The requirement keyword is mapped into a requirement IPC classification, and the requirement IPC classification and the result IPC classification are calculated for an IPC classification distance in an IPC classification tree.

[0008] In combination with the first aspect, in a fourth implementation manner of the first aspect of the application, the shortest relationship path is queried according to the requirement industry label and the result industry label, the path hop count is obtained and divided by a preset maximum path length to obtain the industry correlation distance, including: Cypher shortest path query statement is constructed with the demand industry label as a starting node and the achievement industry label as a target node, and is sent to a Neo4j database through a graph query interface to execute a query and obtain a query result; A path hop number is obtained by counting the number of relationship edges in the path, and an industry association distance is obtained by dividing the path hop number by a preset maximum path length.

[0009] In a fifth implementation manner of the first aspect, the method further includes: A pair of samples of completed conversion are extracted from a historical conversion case database, and sample technology semantic distances, sample industry association distances, and sample IPC classification distances of each pair of samples are used to form a training data set; An inverse matrix of a product of a design matrix transpose and a design matrix is calculated based on the training data set, and then multiplied by the design matrix transpose and a target vector to obtain a first regression coefficient corresponding to the sample technology semantic distance, a second regression coefficient corresponding to the sample industry association distance, and a third regression coefficient corresponding to the sample IPC classification distance; The first regression coefficient, the second regression coefficient, and the third regression coefficient are normalized to obtain a technology semantic weight, an industry association weight, and an IPC classification weight.

[0010] In a sixth implementation manner of the first aspect, the cosine distance, the industry association distance, and the IPC classification distance are weighted to obtain a comprehensive matching score corresponding to each achievement, a target achievement list is filtered based on the comprehensive matching score, and a recommendation result is returned, including: The cosine distance is multiplied by the technology semantic weight, the industry association distance is multiplied by the industry association weight, and the IPC classification distance is multiplied by the IPC classification weight to obtain a comprehensive matching score corresponding to each achievement; An achievement identifier of each achievement in the achievement database and the corresponding comprehensive matching score are combined to form an achievement matching record, and all achievement matching records are summarized to obtain all achievement matching records; The all achievement matching records are sorted in descending order to obtain an achievement matching record set; Achievement matching records with a comprehensive matching score greater than or equal to a preset threshold in the achievement matching record set are filtered to obtain a target achievement list; The recommendation result is returned according to the target achievement list.

[0011] In a seventh implementation form of the first aspect, the returning a recommendation result according to the target achievement list comprises: extracting top-K achievement matching records from the target achievement list, obtaining a path node sequence corresponding to each achievement by querying an association path between an achievement query demand industry label and an achievement industry label in each achievement matching record; encapsulating a ranking number, a comprehensive matching score, a cosine distance, an industry association distance, an IPC classification distance and the path node sequence of each achievement into a recommendation record object, and constructing a JSON format response body by aggregating all recommendation record objects; sending the JSON format response body to a calling party through a RESTful interface of an HTTP GET method to obtain the recommendation result.

[0012] In an eighth implementation form of the first aspect, the extracting top-K achievement matching records from the target achievement list, obtaining a path node sequence corresponding to each achievement by querying an association path between an achievement query demand industry label and an achievement industry label in each achievement matching record comprises: extracting top-K achievement matching records from a starting position of the target achievement list to obtain a top-K achievement information set; constructing a query request parameter by using the demand industry label and an achievement industry label of a corresponding achievement for each achievement in the top-K achievement information set; sending the query request parameter to a Neo4j database to perform an association path query to obtain a graph query result corresponding to each achievement; extracting a path node sequence between the demand industry label and the achievement industry label from the graph query result corresponding to each achievement.

[0013] In a second aspect, the present application provides a scientific and technological achievement intelligent matching system based on multi-dimensional semantic analysis and a knowledge graph, which comprises: a demand extraction module configured to extract a technical parameter dictionary, a demand keyword and a demand industry label from a demand text, and construct demand structured data; a semantic coding module configured to perform word segmentation and coding on the demand structured data to obtain a demand semantic vector, and construct an achievement semantic vector of each achievement in an achievement database; a distance calculation module configured to calculate a cosine distance between the demand semantic vector and the achievement semantic vector, an industry association distance between the demand industry label and an achievement industry label, and an IPC classification distance between a demand keyword and an achievement IPC classification; The matching recommendation module is configured to calculate the cosine distance, the industry association distance and the IPC classification distance by weighting, obtain a comprehensive matching score corresponding to each achievement, filter a target achievement list based on the comprehensive matching score, and return a recommendation result.

[0014] In the technical solution provided by the present application, the micro-service decoupling of BERT semantic encoding, Neo4j graph query and IPC classification calculation is realized through the RESTful interface architecture, the independent deployment, horizontal expansion and parallel calling of each calculation module are supported, and the matching throughput of a large-scale achievement library is improved. The semantic vectors of the demand text and the achievement text are extracted by using the BERT-base model, the deep semantic association recognition across different glossaries is realized by calculating the cosine distance in the vector space, and the matching failure problem caused by the difference in expression in the traditional keyword matching is solved. A fusion system of technical semantic distance, industry association distance and IPC classification distance is constructed, the industry association path hop count is queried through the Neo4j graph database, the IPC classification tree level difference is calculated, and the matching degree of the demand and the achievement in the technical content, the industry application and the patent protection is comprehensively evaluated. A weight optimization mechanism driven by a historical conversion case database is adopted to solve the weight combination that makes the conversion time the shortest, and the technical feasibility, the industry adaptability and the patent matching degree of the recommendation result are ensured, thereby providing accurate and efficient intelligent matching services for the conversion of scientific and technological achievements. BRIEF DESCRIPTION OF DRAWINGS

[0015] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0016] Figure 1 The flowchart of the intelligent matching method of scientific and technological achievements based on multi-dimensional semantic analysis and knowledge graph provided by the embodiments of the present application is shown. Figure 2 The structural schematic diagram of the intelligent matching system of scientific and technological achievements based on multi-dimensional semantic analysis and knowledge graph provided by the embodiments of the present application is shown. DETAILED DESCRIPTION

[0017] The technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are some embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor belong to the protection scope of the present application.

[0018] The flowcharts shown in the drawings are merely illustrative and do not necessarily include all contents and operations / steps, nor are they necessarily executed in the order described. For example, some operations / steps can be further decomposed, combined or partially merged, so the actual execution order can be changed based on actual situations.

[0019] It should also be understood that the terms used in this specification of the present application are only for the purpose of describing particular embodiments and are not intended to limit the present application. As used in this specification and the appended claims, the singular forms "a," "an" and "the" are intended to include plural forms unless the context clearly dictates otherwise.

[0020] It should be further understood that the term "and / or" used in this specification of the present application means any combination of one or more of the associated listed items and all possible combinations, and includes these combinations.

[0021] Some embodiments of the present application will be described in detail below with reference to the accompanying drawings. The following embodiments and features in the embodiments can be combined with each other without conflict.

[0022] Please refer to Figure 1 , Figure 1 The flowchart of the intelligent matching method for scientific and technological achievements based on multi-dimensional semantic analysis and knowledge graph provided by the embodiments of the present application is shown in Figure 1 The intelligent matching method for scientific and technological achievements based on multi-dimensional semantic analysis and knowledge graph provided by the embodiments of the present application includes: Step S100, extracting technical parameter dictionary, demand keyword and demand industry label in demand text, and constructing demand structured data; Specifically, the original requirement text input by the user through the Web end is received, which contains technical indicators, application scenarios, and natural language description of functional requirements. The requirement text is preprocessed in a unified format, and the technical parameter items are identified through a regular expression engine and a numerical normalization module. For example, when the text contains "high temperature resistance 300℃", "tensile strength not less than 800MPa" or "density less than 2.2 grams per cubic centimeter", the numerical fields and physical units are separated, normalized, converted and checked. Equivalent parameters in different expressions are stored in the technical parameter dictionary in the form of key-value pairs, such as {"temperature": 300, "unit": "℃"} and {"strength": 800, "unit": "MPa"}. Structured parameter data with numerical explicitness and unit specification are constructed. Perform keyword recognition process based on word vector model or TF-IDF algorithm, identify representative high-frequency professional terms as demand keywords, such as identifying "aviation", "engine" and "blade" from "applied to aircraft engine blade". Based on the pre-set industry classification mapping table, the keywords are mapped and matched with the field classification to determine the industry label to which the demand belongs, such as "aerospace manufacturing" or "high temperature structural material". The original requirement text, extracted technical parameter dictionary, identified demand keywords and mapped demand industry label are packaged into JSON structured request body in a unified format, such as {"demand_text": original text, "demand_params": parameter dictionary, "demand_industry": industry label}, and an API authentication token is attached as a permission credential. The HTTP POST method is used to send to the BERT encoding service interface deployed in the backend, such as "http: / / server / api / bert / encode". The interface processes the text information to generate semantic vectors after receiving the request, and obtains the demand structured data.

[0023] Step S200, the demand structured data is segmented and encoded to obtain the demand semantic vector, and the achievement semantic vector of each achievement in the achievement database is constructed; Specifically, the tokenizer module provided by the BERT-base model is called for the original demand text field in the structured data of the demand, and the WordPiece segmentation strategy is adopted to segment the text content at the sub-word level. The WordPiece segmentation strategy can effectively process the semantic segmentation of professional terms, out-of-vocabulary words and compound words, so that the model input has sufficient lexical granularity. After segmentation, special marks [CLS] and [SEP] are added at the beginning and end of the generated token sequence, respectively. The [CLS] mark is used to identify the semantic representation position of the whole sentence, and the [SEP] mark is used to indicate the sentence boundary, thereby forming a demand token sequence in a standard format. The demand token sequence is input into the embedding layer of the BERT model, which is composed of three parts: token embedding, segment embedding and position embedding, each of which represents word meaning, sentence fragment position and word position offset in the sequence. The three kinds of information are combined into a unified token representation matrix by vector addition, and the dimension of the matrix is [sequence length x 768]. The token representation matrix is input into the Transformer encoder in the BERT-base model. The encoder is composed of 12 layers of stacked Transformers, each of which contains 12 parallel self-attention heads and a feedforward neural network component. In the encoding stage, the Query, Key and Value matrices are calculated through the multi-head self-attention mechanism, and then the attention weight is obtained and the semantic representation of each token in the context is updated. The vector at the [CLS] position in the output hidden state matrix is extracted as the overall semantic vector of the demand text. At the same time, according to the same process as the demand text, tokenizer segmentation, embedding layer mapping, Transformer multi-layer encoding and CLS vector extraction operations are performed on each achievement description text in the scientific and technological achievement database, and the semantic vector obtained for each achievement is associated with its original identifier to form an achievement semantic vector index structure. The above BERT encoding process is uniformly called through the RESTful interface deployed in the backend, so that the extraction logic of the demand semantic vector and the achievement semantic vector remains consistent, and supports independent deployment and horizontal expansion of the model under high-concurrency matching requests.

[0024] Step S300, calculate the cosine distance of the demand semantic vector and the achievement semantic vector, calculate the industry association distance of the demand industry label and the achievement industry label, and calculate the IPC classification distance of the demand keyword mapping IPC classification and the achievement IPC classification; Specifically, the system receives a 768-dimensional demand semantic vector and an achievement semantic vector, calls a RESTful interface to send a POST request to a technology distance calculation module and carry the two vectors, calculates the similarity between the two vectors through a standard cosine distance formula, calculates the dot product and the modulus product of the two vectors respectively, takes the quotient value as the cosine similarity, and subtracts the cosine similarity from 1 to obtain the cosine distance. The cosine distance takes a value ranging from 0 to 1, and the closer to 0, the higher the semantic similarity. Based on the industry label relationship graph structure recorded in the knowledge graph, a graph query language Cypher is used to initiate a shortest path query in a graph database. The shortest directed path between the demand industry label node and the achievement industry label node is searched, and the number of edges contained in the path is returned as the path hop number. Then, the path hop number is divided by the preset maximum path length (such as 5 hops) to obtain a normalized industry association distance value, and the result is located between 0 and 1. The fewer the hops, the stronger the industry association. The IPC classification number mapped by the demand keyword and the IPC classification number annotated by the achievement itself are compared in structure level. An IPC classification structure interface is called to locate the positions of the two classification numbers in an IPC eight-level hierarchical tree, find the common ancestor node, and calculate the level difference value from the ancestor node. The sum of the two level differences is divided by the maximum number of layers of the IPC classification tree (such as 8 layers) to obtain the IPC classification distance, which reflects the closeness of the two technical topics in the classification system.

[0025] Step S400, the cosine distance, industry association distance and IPC classification distance are weighted and calculated to obtain a comprehensive matching score corresponding to each achievement. Based on the comprehensive matching score, a target achievement list is screened and a recommendation result is returned.

[0026] Specifically, multiply the technical semantic cosine distance of each result by the technical semantic weight value ω1, such as 0.5, multiply the industry correlation distance by the industry weight ω2, such as 0.3, multiply the IPC classification distance by the IPC weight ω3, such as 0.2, add the three products to obtain the weighted distance value of the current result, and then subtract the weighted distance value from 1 to obtain the comprehensive matching score. The closer the score value is to 1, the higher the matching degree with the target demand. Structurally bind the comprehensive matching score to the corresponding result identifier to generate a result matching record containing {result ID, comprehensive matching score}, and in this way, all result entries in the result database are traversed to construct a result matching record set. All result records in the result matching record set are sorted in descending order according to the comprehensive matching score field to obtain a result matching record list with clear priority. The sorting process uses quicksort or heapsort algorithm to improve processing efficiency. Set a preset score threshold, such as 0.6, and filter out all result matching records with a matching score greater than or equal to the score threshold in the sorting result to form a target result list. Assemble all matching items in the target result list in descending order to form a recommendation result, which is packaged as a structured JSON response body containing fields such as result identifier, matching score, distance decomposition value, industry path, and key technology parameters, and returned to the front-end caller through a RESTful interface.

[0027] In a specific embodiment, the process of step S100 can specifically include the following steps: Extract the numerical value and unit of the technical parameter from the demand text to obtain a technical parameter dictionary; Perform keyword recognition on the demand text to obtain demand keywords, and map the demand keywords to demand industry labels according to a preset industry classification mapping table; Package the demand text, technical parameter dictionary, and demand industry label into a request body structure, and send it to the BERT encoding interface through the HTTP POST method with API authentication Token to obtain structured data.

[0028] Specifically, a unified preprocessing process is performed on the user input original natural language demand text, and regular expression recognition patterns, unit standardization library, and numerical value extraction algorithm are used to sequentially traverse all character fragments in the demand text, and pattern matching is performed on expression structures containing numerical value type technical indicators, such as "working temperature not less than 300℃", "voltage resistance capacity should reach 50 megapascal", "density less than 2.1g / cm 3The phrases such as "300 degrees Celsius", "50 MPa", and "2.1 g / cm3" are located by the lexical window, the Arabic numerals, Chinese numerals, and unit identifiers contained therein are standardized into a unified format of numerical-value pairs according to the unit conversion table, and a technical parameter dictionary in the form of key-value is constructed, such as {"temperature": 300, "unit": "degrees Celsius"}, {"pressure": 50, "unit": "MPa"}, and {"density": 2.1, "unit": "g / cm3"}. Based on the keyword identification module, the module analyzes the noun phrases, professional terms, and verb-object structures in the text by combining the TF-IDF model, the domain dictionary, and the part-of-speech tagging mechanism, and filters out representative high-frequency concepts as demand keywords, such as "aviation", "engine", "blade", "high-temperature resistant", and "lightweight material" extracted from "high-temperature resistant lightweight material for aviation engine blades". According to the industry classification mapping table, the mapping table associates typical keywords or keyword combinations with standardized industry labels, such as mapping the keyword set containing the combination of "engine" and "aviation" to the "aerospace manufacturing" industry label. After the keyword set hits the rule, the corresponding demand industry label is generated to identify the technical field category to which the demand belongs. The original demand text field, the technical parameter dictionary parsed, and the industry label are assembled into a structured request body according to the predefined JSON Schema standard, and the HTTP request header includes the API identity authentication Token for security verification. The structured data packet is sent to the BERT encoding interface (such as http: / / server / api / bert / encode) deployed in the backend through the HTTP POST method, and the BERT encoding interface completes the deep semantic modeling of the text content and returns the semantic vector representation while retaining the structured field.

[0029] In a specific embodiment, the process of performing step S200 can specifically include the following steps: The tokenizer of the BERT-base model is called to perform WordPiece segmentation on the demand text in the demand structured data, to split into a token sequence and add CLS and SEP markers at the beginning and end of the sequence, to obtain a demand token sequence; Each token in the demand token sequence is mapped to a token matrix through token embedding, segment embedding, and position embedding of the embedding layer; The token matrix is subjected to multi-head self-attention calculation to obtain a target output matrix, and a demand semantic vector corresponding to the CLS marker is extracted from the target output matrix; The BERT encoding interface is called to generate an achievement semantic vector according to each achievement in the achievement database.

[0030] Specifically, for the text field in the structured data of the demand, i.e. the original natural language demand description, the tokenizer component built in the BERT-base model is called, the tokenizer component adopts the WordPiece algorithm for sub-word level segmentation, divides the complete sentence into several token segments with semantic boundaries, inserts the [CLS] mark in front of the generated token sequence to represent the overall semantic representation entrance, and inserts the [SEP] mark at the end to represent the semantic closure, thereby forming an input sequence. The token sequence is input into the embedding layer for embedding vector mapping, the embedding layer is composed of three sub-modules, namely token embedding for representing token semantics, segment embedding for identifying sentence boundaries, and position embedding for representing the absolute position index of the token in the sequence. The three parts of vectors are added bit by bit to generate a token embedding matrix, which has a dimension of [sequence length x 768]. The token embedding matrix is input into the 12-layer Transformer encoding structure of the BERT-base model. Each layer of the Transformer module contains 12 independent multi-head self-attention calculation channels. By constructing Query, Key and Value matrices, the attention weight between tokens is calculated to realize the modeling of the context dependence between tokens and the dynamic semantic aggregation. After 12 layers of recursive encoding, the 768-dimensional vector corresponding to the CLS position in the final output matrix is extracted as the overall semantic vector representation of the demand text. At the same time of completing the vectorization of the demand text, all the matched scientific and technological achievement description texts in the achievement database are traversed, and the tokenizer segmentation, embedding mapping, multi-head attention encoding and CLS vector extraction processes described above are repeated for each achievement text to obtain the independent semantic vector corresponding to each achievement.

[0031] In a specific embodiment, the process of performing step S300 can specifically include the following steps: Calculate the cosine distance between the demand semantic vector and the achievement semantic vector; According to the demand industry label and the achievement industry label, query the shortest relationship path, obtain the path hop count and divide it by the preset maximum path length to obtain the industry association distance; Map the demand keywords to the demand IPC classification, and calculate the IPC classification distance between the demand IPC classification and the achievement IPC classification in the IPC classification tree.

[0032] Specifically, the demand semantic vector and the achievement semantic vector are input into the cosine similarity calculation module as a vector pair, the cosine value of the included angle is calculated according to the formula Sim(V1, V2) = (V1·V2) / (||V1|| × ||V2||), and the cosine distance after normalization is obtained by subtracting the cosine similarity value from 1, and the smaller the cosine distance is, the higher the semantic similarity degree is. The industry correlation query process based on the graph database is started, the demand industry label and the achievement industry label are extracted from the structured request body, and the two are submitted to the Neo4j graph database as the starting node and the ending node respectively, and the Cypher statement such as MATCH path=shortestPath((n:Industry)-[*]-(m:Industry)) RETURN length(path) is executed to obtain the shortest path hop count between the industry labels, and the shortest path hop count value is divided by the preset maximum industry graph path length (for example, 5) to obtain the industry correlation distance, and the smaller the distance value is, the closer the industry relationship is. The technical category distance between the IPC classifications is evaluated, the IPC classification number of the demand text is obtained by executing the IPC mapping rule (for example, “aviation”, “engine” and “turbine blade” are mapped to F01D) on the keywords in the demand text, and the IPC main class (for example, C04B35 / 80) given in the achievement metadata is compared in structure, the nearest common ancestor node of the two in the IPC classification tree is found, and the level difference of the two to the ancestor node is counted respectively, and the total difference is divided by the maximum level of the IPC tree (for example, 8) to obtain the IPC classification distance, which measures the closeness of the two technical fields in the patent knowledge structure. The three types of distance calculation modules are deployed in the form of RESTful microservices, which can support parallel input and independent calculation, so that there is no delay bottleneck caused by serial dependence in large-scale matching tasks.

[0033] In a specific embodiment, the process of querying the shortest relationship path according to the demand industry label and the achievement industry label, obtaining the path hop count and dividing it by the preset maximum path length to obtain the industry correlation distance can specifically include the following steps: The Cypher shortest path query statement is constructed with the demand industry label as the starting node and the achievement industry label as the target node, is sent to the Neo4j database through the graph query interface to execute the query, and the query result is obtained. The relationship path node sequence from the starting node to the target node is extracted from the query result, the number of relationship edges in the path is counted to obtain the path hop count, and the path hop count is divided by the preset maximum path length to obtain the industry correlation distance.

[0034] Specifically, based on the industry label information parsed from the demand structured data, the demand sub-industry label and the achievement sub-industry label are extracted as the starting point and the end point nodes of the shortest path query in the knowledge graph. In the graph database query statement construction phase, according to the Cypher query language supported by the Neo4j database, the industry label name is embedded into the query template as the attribute of the graph node to build a standardized statement, the specific form is: MATCH path = shortestPath ((n: Industry {name: 'demand label'})-[*]-(m: Industry {name: 'achievement label'})) RETURN path, wherein Industry is the label to which the industry entity in the graph database belongs, name is its unique identification attribute, and [*] represents performing path search on edges of any type and in any direction. After the construction is completed, the statement is sent to the graph query RESTful interface deployed in the backend through encapsulation in the HTTP POST request body, and an interface authentication token is attached to ensure access permission. After the Neo4j database successfully parses and executes the query, the returned data structure contains a shortest relationship path from the starting node to the target node. The names, types and connection relationships of the nodes in the path are parsed and extracted from the query result, and a complete path node sequence is constructed in the form of an array. The number of edges contained in the shortest relationship path, that is, the number of direct relationships experienced in the process from the first node to the last node, is counted, and this value is defined as the path hop count. A global and unified maximum industry hop count limit is set, for example, the value is 5, the path hop count is divided by the maximum hop count value, and a standardized industry association distance is obtained, the distance value range is between 0 and 1, and the smaller the distance, the closer the topological association of the two industry labels in the knowledge graph.

[0035] In a specific embodiment, the technology achievement intelligent matching method based on multi-dimensional semantic analysis and knowledge graph further includes the following steps: Extract the completed conversion paired samples from the historical conversion case database, and construct the training data set from the sample technology semantic distance, sample industry association distance and sample IPC classification distance of each paired sample; Based on the training data set, calculate the inverse matrix of the design matrix transpose and the design matrix product, and then multiply the design matrix transpose and the target vector to obtain the first regression coefficient corresponding to the sample technology semantic distance, the second regression coefficient corresponding to the sample industry association distance and the third regression coefficient corresponding to the sample IPC classification distance; The first regression coefficient, the second regression coefficient and the third regression coefficient are normalized respectively to obtain the technology semantic weight, the industry association weight and the IPC classification weight.

[0036] Specifically, the demand-outcome pair records of successful scientific and technological achievement transformation are extracted in batches from the historical transformation case database through the RESTful interface, each record contains the three-dimensional distance features calculated in the matching process, namely the technical semantic distance D1, the industry correlation distance D2 and the IPC classification distance D3, and also contains the actual transformation effect index, such as the transformation time T from recommendation to signing of agreement. Each record is arranged as a set of training samples to construct a sample matrix X, where each row is [D1, D2, D3], and a corresponding target vector Y is constructed, where each item is the T value corresponding to the sample. According to the principle of least squares, the linear regression solving process is constructed, the transpose X t of the design matrix X is calculated t , the inverse matrix of X t is solved -1 , the inverse matrix is multiplied by X t , and multiplied by the target vector Y to form (X t ·X) -1 ·X t ·Y, and the regression coefficient vector B=[β1, β2, β3] is obtained, where β1 represents the sensitivity of the technical semantic distance to the transformation time, β2 represents the influence weight of the industry correlation distance, and β3 represents the correlation strength of the IPC classification structure. The three regression coefficients are normalized in proportion, that is, the sum S=β1+β2+β3 of β1, β2 and β3 is calculated, and each coefficient is divided by the sum to form the normalized weight vector ω=[β1 / S,β2 / S, β3 / S].

[0037] In a specific embodiment, the process of performing step S400 can specifically include the following steps: The cosine distance is multiplied by the technical semantic weight, the industry correlation distance is multiplied by the industry correlation weight, and the IPC classification distance is multiplied by the IPC classification weight to obtain the comprehensive matching score corresponding to each outcome; The outcome identifier and the corresponding comprehensive matching score of each outcome in the outcome database are combined to construct an outcome matching record, and all outcome matching records are summarized to obtain all outcome matching records; The all outcome matching records are sorted in descending order to obtain an outcome matching record set; The outcome matching records with a comprehensive matching score greater than or equal to a preset threshold value in the outcome matching record set are screened to obtain a target outcome list; The recommendation result is returned according to the target outcome list.

[0038] Specifically, the weight coefficients obtained by historical regression optimization are called in sequence for the technical semantic cosine distance, the industry correlation distance, and the IPC classification distance, respectively, numerical combination is performed according to the weighted fusion formula, the cosine distance D1 corresponding to a certain achievement is multiplied by the technical semantic weight ω1, the industry distance D2 is multiplied by the industry weight ω2, and the IPC classification distance D3 is multiplied by the IPC weight ω3, and the sum of the three products is recorded as the weighted distance value D_total = ω1×D1+ ω2×D2+ ω3×D3, and then the system calculates the matching score Smatch = 1 - D_total to convert the distance index into a standardized similarity index, and the closer the Smatch value is to 1, the closer the achievement is in the three dimensions to the demand. The unique identifier of the current achievement and its corresponding matching score Smatch are packaged as an achievement matching record, and all achievement records in the achievement database are processed in the same way to generate corresponding matching records one by one and form a matching record set. After the matching record set is constructed, a quicksort algorithm or a stable heap sorting mechanism is called to sort all achievement matching records in descending order according to the score field value, to construct an ordered achievement set with clear matching priority, and a score lower limit threshold (such as 0.6) predefined by the system or configurable by the user is set to perform a filtering operation on the sorted set, and all achievement records with a matching score not lower than the score lower limit threshold are retained as the target achievement set. The target achievement set is renumbered according to the rank field and encapsulated as a recommendation result output by calling the RESTful response interface, which includes achievement ID and score, with three types of distance value details, graph path information, core technology parameters, and patent status fields, and is returned to the front-end caller in a structured JSON format.

[0039] In a specific embodiment, the process of returning the recommendation result according to the target achievement list can specifically include the following steps: Extract the top K achievement matching records from the target achievement list, and query the industry label association path of each achievement matching record to obtain the path node sequence corresponding to each achievement. Encapsulate the ranking number, comprehensive matching score, cosine distance, industry correlation distance, IPC classification distance, and path node sequence of each achievement into a recommendation record object, and aggregate all recommendation record objects to construct a JSON format response body; Send the JSON format response body to the caller through the RESTful interface of the HTTP GET method to obtain the recommendation result.

[0040] Specifically, the comprehensive matching scores are ranked in descending order and the top K matching records are extracted. Based on the extracted demand industry labels and the industry labels of each achievement, a shortest path query statement is constructed, and the Cypher statement is sent to the deployed graph query interface in turn, such as MATCH path=shortestPath((n:Industry{name:'demand industry'})-[*]-(m:Industry{name:'achievement industry'})) RETURN nodes(path). The query is executed through the Neo4j database to obtain the graph path node sequence from the demand industry to the achievement industry. The path is parsed into an ordered node array, representing the semantic migration relationship chain between industries. For each of the top K achievements, a recommendation record object is constructed, which includes the ranking position of the achievement in the recommendation result (such as rank: 1~K), the comprehensive matching score Smatch, the three-dimensional distance decomposition value D_tech, D_industry, D_ipc, and the industry path node sequence obtained through the graph database, and the information is organized using a unified field structure. After constructing K recommendation record objects, all objects are merged into a recommendation result list top_matches, and other global information such as request_id, matching threshold, target achievement quantity, total response time, etc. are combined into a JSON structure. The GET method of the RESTful interface is called, the JSON response body is sent to the front-end caller, the interface response content is set to Content-Type: application / json, and the HTTP status code 200 is returned to indicate that the request is successful. The front-end system receives the recommendation result, parses the JSON structure, and visualizes each recommendation record through the page component to display to the user.

[0041] In a specific embodiment, the execution step extracts the top K matching records from the target achievement list, and the process of querying the associated path of the demand industry label to the achievement industry label for each achievement in the achievement matching record to obtain the path node sequence corresponding to each achievement can specifically include the following steps: From the starting position of the target achievement list, the first K achievement matching records are sequentially extracted to obtain a set of top K achievement information; For each achievement in the set of top K achievement information, the demand industry label and the achievement industry label corresponding to the achievement are constructed as query request parameters; Send the query request parameters to the Neo4j database to execute the associated path query and obtain the graph query result corresponding to each achievement; From the graph query result corresponding to each achievement, the path node sequence between the demand industry label and the achievement industry label is parsed and extracted.

[0042] Specifically, based on the target achievement matching record, the matching scores are sorted in descending order and the range is determined, the K value (such as K = 10) is set, the first K achievements are sequentially intercepted from the starting position of the sorted list to form a top K achievement information set, and each achievement in the set contains an achievement unique identifier, a comprehensive matching score, three types of distance values and an industry label. For each achievement, the industry label of the current demand is extracted from the structured demand data, the industry label and the industry label of the achievement are used as the start and end parameters of the path query, and they are packaged into a graph query request format, such as the parameter {"source": "aerospace manufacturing", "target": "high-temperature structural material application"}. The graph path query service is called through the RESTful interface, and the service initiates a query to the Neo4j graph database through a Cypher statement in the back end, for example: MATCH path = shortestPath ((n: Industry {name:'source'})- [*]-(m: Industry {name: 'target'})) RETURN nodes (path), the database returns a shortest path from source to target according to the graph structure between industry nodes. The query result is parsed, and the node object array is extracted therefrom to form a node name sequence in the order of the path, for example ["aerospace manufacturing", "high-temperature material", "ceramic matrix composite", "high-temperature structural material application"], and the node name sequence is the semantic path trajectory between the demand industry and the achievement industry, indicating the industry field migration or adaptation relationship. For the top K achievements, each repeats the above query and path extraction process to obtain K industry path node sequences.

[0043] Please refer to Figure 2 , Figure 2 The structure schematic diagram of the intelligent matching system 200 of scientific and technological achievements based on multi-dimensional semantic analysis and knowledge graph provided by the embodiments of the present application is shown in Figure 2 The intelligent matching system 200 of scientific and technological achievements based on multi-dimensional semantic analysis and knowledge graph includes: A demand extraction module 210 is configured to extract technical parameter dictionaries, demand keywords and demand industry labels from demand text, and construct structured demand data. A semantic coding module 220 is configured to perform word segmentation and coding on the structured demand data to obtain a demand semantic vector, and construct an achievement semantic vector of each achievement in an achievement database. A distance calculation module 230 is configured to calculate the cosine distance between the demand semantic vector and the achievement semantic vector, the industry association distance between the demand industry label and the achievement industry label, and the IPC classification distance between the demand keyword and the achievement IPC classification. The matching recommendation module 240 is configured to perform weighted calculation on the cosine distance, the industry association distance and the IPC classification distance, to obtain a comprehensive matching score corresponding to each achievement, filter a target achievement list based on the comprehensive matching score, and return a recommendation result.

[0044] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described system, system and unit can refer to the corresponding processes in the foregoing method embodiments, which will not be described here.

[0045] The integrated unit, if realized in the form of a software function unit and sold or used as an independent product, can be stored in a computer-readable storage medium. Based on such understanding, the technical solutions of the present application, essentially or in other words, the part that contributes to the prior art or the whole or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes a number of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The foregoing storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various program code storage media.

[0046] The above-described embodiments are only used to illustrate the technical solutions of the present application, rather than limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for intelligent matching of scientific and technological achievements based on multi-dimensional semantic analysis and knowledge graphs, characterized in that, include: Extract the technical parameter dictionary, requirement keywords, and requirement industry tags from the requirement text, and construct the requirement structured data; The structured demand data is segmented and encoded to obtain demand semantic vectors, and the result semantic vectors of each result in the result database are constructed. Calculate the cosine distance between the demand semantic vector and the result semantic vector, calculate the industry association distance between the demand industry tag and the result industry tag, map the demand keywords to IPC classification and result IPC classification, and calculate the IPC classification distance. The cosine distance, the industry association distance, and the IPC classification distance are weighted and calculated to obtain a comprehensive matching score for each result. Based on the comprehensive matching score, a list of target results is filtered and a recommendation result is returned.

2. The intelligent matching method for scientific and technological achievements based on multi-dimensional semantic analysis and knowledge graph as described in claim 1, characterized in that, The process involves extracting the technical parameter dictionary, keyword, and industry tags from the requirement text and constructing structured requirement data, including: Extract the numerical values ​​and units of the technical parameters from the requirement text to obtain a technical parameter dictionary; Keyword recognition is performed on the demand text to obtain demand keywords, and the demand keywords are mapped to demand industry tags according to a preset industry classification mapping table; The requirement text, the technical parameter dictionary, and the requirement industry tag are encapsulated into a request body structure, and sent to the BERT encoding interface via the HTTP POST method carrying the API authentication token to obtain the requirement structured data.

3. The intelligent matching method for scientific and technological achievements based on multi-dimensional semantic analysis and knowledge graph as described in claim 1, characterized in that, The process of segmenting and encoding the structured demand data to obtain demand semantic vectors, and constructing the result semantic vector for each result in the result database, includes: The BERT-base model's tokenizer is invoked to perform WordPiece segmentation on the demand text in the demand structured data, dividing it into a token sequence and adding CLS and SEP tags at the beginning and end of the sequence to obtain the demand token sequence; Each token in the required token sequence is mapped to a token matrix through token embedding, segment embedding, and position embedding in the embedding layer; Perform multi-head self-attention computation on the token matrix to obtain the target output matrix, and extract the demand semantic vector corresponding to the CLS tag from the target output matrix; Call the BERT encoding interface to generate semantic vectors for each result in the results database.

4. The intelligent matching method for scientific and technological achievements based on multi-dimensional semantic analysis and knowledge graph as described in claim 1, characterized in that, The calculation of the cosine distance between the demand semantic vector and the result semantic vector, the calculation of the industry association distance between the demand industry tag and the result industry tag, and the mapping of demand keywords to IPC categories and result IPC categories to calculate the IPC category distance include: Calculate the cosine distance between the demand semantic vector and the outcome semantic vector; Based on the industry tags of the demand and the industry tags of the results, the shortest relationship path is queried, the number of path jumps is obtained and divided by the preset maximum path length to obtain the industry association distance; The demand keywords are mapped to demand IPC categories, and the IPC category distance between the demand IPC category and the result IPC category is calculated in the IPC category tree.

5. The intelligent matching method for scientific and technological achievements based on multi-dimensional semantic analysis and knowledge graph as described in claim 4, characterized in that, The step of querying the shortest relationship path based on the demand industry tag and the result industry tag, obtaining the path hop count and dividing it by the preset maximum path length to obtain the industry association distance includes: Using the demand industry label as the starting node and the result industry label as the target node, construct a Cypher shortest path query statement, send it to the Neo4j database through the graph query interface to execute the query and obtain the query results; Extract the sequence of relational path nodes from the starting node to the target node from the query results, count the number of relational edges in the path to obtain the path hop count, and divide the path hop count by the preset maximum path length to obtain the industry association distance.

6. The intelligent matching method for scientific and technological achievements based on multi-dimensional semantic analysis and knowledge graph as described in claim 4, characterized in that, The intelligent matching method for scientific and technological achievements based on multi-dimensional semantic analysis and knowledge graphs also includes: Extract paired samples that have completed conversions from the historical conversion case database, and construct a training dataset by taking the sample technology semantic distance, sample industry association distance, and sample IPC classification distance of each pair of paired samples; Based on the training dataset, calculate the inverse matrix of the product of the transpose of the design matrix and the design matrix, and then multiply it by the transpose of the design matrix and the target vector to obtain the first regression coefficient corresponding to the technical semantic distance of the sample, the second regression coefficient corresponding to the industry association distance of the sample, and the third regression coefficient corresponding to the IPC classification distance of the sample. The first regression coefficient, the second regression coefficient, and the third regression coefficient are normalized respectively to obtain the technical semantic weight, the industry association weight, and the IPC classification weight.

7. The intelligent matching method for scientific and technological achievements based on multi-dimensional semantic analysis and knowledge graph as described in claim 6, characterized in that, The process involves weighting the cosine distance, industry association distance, and IPC classification distance to obtain a comprehensive matching score for each result. Based on this comprehensive matching score, a list of target results is filtered and recommended results are returned, including: Multiply the cosine distance by the technical semantic weight, the industry association distance by the industry association weight, and the IPC classification distance by the IPC classification weight to obtain the comprehensive matching score for each result; The result identifier and corresponding comprehensive matching score of each result in the results database are combined to construct a result matching record. All result matching records are then aggregated to obtain the complete result matching records. Sort all the result matching records in descending order to obtain a set of result matching records; Filter the result matching records in the result matching record set whose comprehensive matching score is greater than or equal to a preset threshold to obtain the target result list; Return recommended results based on the list of target outcomes.

8. The intelligent matching method for scientific and technological achievements based on multi-dimensional semantic analysis and knowledge graph as described in claim 7, characterized in that, The step of returning recommendation results based on the target outcome list includes: Extract the top K matching records from the target results list, and obtain the path node sequence corresponding to each result by analyzing the association path from the industry tag of the result query requirement to the industry tag of the result in each matching record. Each result's ranking number, comprehensive matching score, cosine distance, industry association distance, IPC classification distance, and the path node sequence are encapsulated into a recommendation record object. All recommendation record objects are then aggregated to construct a JSON format response body. The JSON-formatted response body is sent to the caller via the HTTP GET method of the RESTful interface to obtain the recommended results.

9. The intelligent matching method for scientific and technological achievements based on multi-dimensional semantic analysis and knowledge graph as described in claim 8, characterized in that, The step involves extracting the top K matching records from the target results list, and for each matching record, obtaining the association path from the industry tag of the query request to the industry tag of the result, resulting in a sequence of path nodes corresponding to each result, including: The first K matching records of the target results are extracted sequentially from the beginning of the target results list to obtain the set of the first K results information; For each result in the set of the top K results, the demand industry tag and the result industry tag of the corresponding result are constructed as query request parameters; The query request parameters are sent to the Neo4j database to perform a related path query, and the graph query results corresponding to each result are obtained. The path node sequence between the demand industry tag and the result industry tag is extracted from the graph query results corresponding to each result.

10. A scientific and technological achievement intelligent matching system based on multi-dimensional semantic analysis and knowledge graph, characterized in that, The method for intelligent matching of scientific and technological achievements based on multi-dimensional semantic analysis and knowledge graphs as described in any one of claims 1-9 includes: The requirement extraction module is used to extract the technical parameter dictionary, requirement keywords and requirement industry tags from the requirement text, and to build the requirement structured data; The semantic encoding module is used to segment and encode the structured demand data to obtain demand semantic vectors and construct the result semantic vector for each result in the result database. The distance calculation module is used to calculate the cosine distance between the demand semantic vector and the result semantic vector, calculate the industry association distance between the demand industry tag and the result industry tag, and map the demand keywords to IPC classification and result IPC classification to calculate the IPC classification distance. The matching and recommendation module is used to perform weighted calculations on the cosine distance, the industry association distance, and the IPC classification distance to obtain a comprehensive matching score for each result, filter the target result list based on the comprehensive matching score, and return the recommendation results.

Images