A method and system for generating a technology list based on knowledge graph

Abstract

The present invention discloses a method for generating a technology list based on a knowledge graph, including: obtaining patent data, inputting the patent data into a pre-built knowledge graph and embedding it into a neural network model to obtain an embedding result of the patent data, that is, an embedding vector of the patent data; using a Canopy clustering algorithm to process the embedding vector of the patent data to obtain K cluster centers, using the K cluster centers as initial centers, using a K-means algorithm to perform clustering processing on the entity embedding vector to obtain K clustering results as K technical fields, and for each of the K technical fields, obtaining all patent texts in each technical field from a patent database. The present invention can solve the technical problems existing in the existing technology list generation method, such as involving many experts, taking a long time, requiring a lot of energy, relying heavily on the knowledge and experience of experts, and lacking quantitative analysis and data support.

Description

Technical Field

[0001] The present invention belongs to the field of knowledge graph technology, and more specifically, relates to a method and system for generating a technology list based on knowledge graph. Background Art

[0002] Technology foresight refers to the forward-looking discovery of technical problems, which facilitates the solution of problems before they occur, or the preparation of solutions in advance. Technology foresight is a strategic management tool commonly used by countries. It is widely used in the formulation of long-term science and technology development strategies and has gradually become a global trend. Dividing all technologies into different fields for discussion, selecting key technologies, and generating a technology list are key steps in the entire technology foresight process.

[0003] The traditional technology list formulation process mainly adopts extensive questionnaire consultation methods such as the Delphi method, relying on the experience of expert knowledge and qualitative methods to determine and rank technologies. The Delphi method, also known as the expert survey method, is essentially a feedback anonymous letter inquiry method. Its general process is to obtain the opinions of experts on the issues to be predicted, sort them out, summarize them, and make statistics, and then anonymously feedback them to the experts, solicit opinions again, concentrate them again, and then feedback them again until a consensus is reached.

[0004] However, the Delphi method and other extensive questionnaire consultation technology list generation methods have some non-negligible defects:

[0005] First, the above method involves many experts, is time-consuming, requires a lot of energy, relies heavily on the knowledge and experience of experts, and lacks quantitative analysis and data support.

[0006] Second, the above methods generally divide technical fields according to established division standards, such as dividing technical fields according to the International Patent Classification (IPC) for analysis. This method of division has the problem of rigid division standards, making it difficult to discover emerging technical fields. The scope and pertinence of the division standards are sometimes difficult to meet actual needs, and cannot change with technological development.

[0007] Third, the description of the technical field in the above method generally only has a single fixed field term, which cannot well describe the new technical field generated by the integration of multiple technical fields, and the description of the technical field cannot change with the development of technology. Summary of the invention

[0008] In view of the above defects or improvement needs of the prior art, the present invention provides a method and system for generating a technology list based on a knowledge graph, the purpose of which is to cluster patent entities, divide the fields by utilizing the structural information in the knowledge graph, obtain the key technical terms of each field by the TF-IDF algorithm, and finally formulate a process method for the technology list of the corresponding field in combination with the evaluation indicators, thereby solving the technical problems existing in the prior art list generation method, such as involving many experts, taking a long time, consuming a lot of energy, relying heavily on the knowledge and experience of experts, lacking quantitative analysis and data support, as well as the technical problems in emerging technology fields that are difficult to discover due to the rigidity of the division standards, and the technical problems that the scope and pertinence of the division standards are sometimes difficult to meet actual needs and cannot change with the development of technology, and the new technology fields generated by the integration of multiple technology fields cannot be well described, and the description of the technology field cannot change with the development of technology.

[0009] To achieve the above object, according to one aspect of the present invention, a method for generating a technology list based on a knowledge graph is provided, comprising the following steps:

[0010] (1) Obtain patent data, input the patent data into a pre-built knowledge graph and embed it into a neural network model to obtain an embedding result of the patent data, i.e., an embedding vector of the patent data;

[0011] (2) Using the Canopy clustering algorithm to process the embedded vectors of the patent data obtained in step (1) to obtain K cluster center points;

[0012] (3) Taking the K cluster centers obtained in step (2) as the initial centers, the entity embedding vectors obtained in step (1) are clustered using the K-means algorithm to obtain K clustering results as K technical fields.

[0013] (4) For each of the K technical fields obtained in step (3), all patent texts in each technical field are obtained from a patent database (e.g., the Derwent patent database), and all patent texts in each technical field are processed using a term frequency-inverse text frequency index TF-IDF algorithm to obtain multiple key technical terms corresponding to the technical field.

[0014] (5) For each of the K technical fields obtained in step (3), obtain its corresponding technology research heat evaluation index and / or technology frontier evaluation index, as well as the value corresponding to the technology research heat evaluation index and / or technology frontier evaluation index, and then sort all K technical fields in descending order according to the values ​​corresponding to the technology research heat evaluation index and / or technology frontier evaluation index corresponding to all K technical fields, and finally generate a mapping relationship between the technology research heat evaluation index and / or technology frontier evaluation index, the sorted K technical fields, and the key technical terms corresponding to each technical field in step (4), thereby forming a technology list and outputting it.

[0015] Preferably, the knowledge graph embedding neural network model is a ConvE model, and is constructed by the following steps:

[0016] (1-1) Obtaining patent datasets from patent databases;

[0017] (1-2) Generate multiple triples based on the patent dataset obtained in step (1-1), all triples constitute a knowledge graph dataset, and each knowledge graph data in the knowledge graph dataset is a triple (e m1 , r mr , e m2 ), where e m1 represents the embedding vector of the head entity in the mth triplet in the knowledge graph dataset, r mr Represents the embedding vector of the relationship in the mth triple in the knowledge graph dataset, e m2 Represents the embedding vector of the tail entity in the mth triple in the knowledge graph dataset, and m∈[1, the total number of triples in the knowledge graph dataset].

[0018] (1-3) Divide the knowledge graph data set obtained in step (1-2) into a training set, a validation set, and a test set. All entities in all knowledge graph data (i.e., triples) in the training set constitute a set E, and all relations in all knowledge graph data (i.e., triples) in the training set constitute a set R;

[0019] (1-4) Randomly initialize the entity set E and the relationship set R corresponding to the training set obtained in step (1-3) to obtain the embedding vector of each entity and each relationship;

[0020] (1-5) Input the embedding vectors of all entities and relationships in the training set obtained in step (1-4) into the knowledge graph embedding neural network model ConvE to obtain multiple prediction scores;

[0021] (1-6) Input all the prediction scores obtained in step (1-5) into the loss function to minimize the loss function L, thereby obtaining a preliminarily trained knowledge graph embedded in the neural network model;

[0022] (1-7) Verifying the knowledge graph embedding neural network model preliminarily trained in step (1-6) according to the verification set obtained in step (1-3) until the loss function L is optimal, thereby obtaining a trained knowledge graph embedding neural network model;

[0023] Preferably, the triple in step (1-2) is composed of head entity-relationship-tail entity;

[0024] Entities include patent title, inventor, applicant, IPC classification number, and country;

[0025] The relationships include applicant-patent name, patent name-IPC classification number, inventor-applicant, inventor-patent name, and applicant-country;

[0026] The triples generated in step (1-2) include (applicant unit-application-patent name), (patent name-belongs to-IPC classification number), (inventor-belongs to-applicant unit), (inventor-application-patent name), and (applicant unit-belongs to-country).

[0027] Preferably, the value range of the embedding vector in step (1-4) is 1 to 1024, preferably 200;

[0028] Each embedding vector satisfies a normal distribution with a mean of 0.

[0029] Preferably, step (1-5) comprises the following sub-steps:

[0030] (1-5-1) For each triple in the training set (e k1 , r kr , e k2 ), the 2D reshape operation is used to transform the triple (e k1 , r kr , e k2 ) is the embedding vector e of the head entity k1 and the embedding vector r of the relation kr Convert into a matrix and concatenate to get the triple (e k1 , r kr , e k2 ) where k∈[1, the total number of triplets in the training set].

[0031] (1-5-2) For each triple in the training set (e k1 , rkr , e k2 ), the triplet (e k1 , r kr , e k2 )The corresponding concatenated matrix is ​​input into the two-dimensional convolutional layer with filter ω in the ConvE model to obtain the tensor corresponding to the triplet.

[0032] (1-5-3) For each triple in the training set (e k1 , r kr , e k2 ), the tensor corresponding to the triplet is flattened into a one-dimensional vector and then input into the fully connected layer with parameter matrix W in the ConvE model to obtain the triplet (e k1 , r kr , e k2 ) corresponds to the output vector.

[0033] (1-5-4) For each triple in the training set, add the vector corresponding to the triple obtained in step (1-5-3) to the embedding vector e of the tail entity in the triple k2 Perform the inner product operation and use the logistic sigmoid function to process the inner product calculation result to obtain the output result as the triple (e1, r r ,e2)’s prediction score;

[0034] Preferably, the prediction score in step (1-5-4) is calculated using the following formula:

[0035]

[0036] where e k1 、e kt and r kr Represent the embedding vectors of the head entity, tail entity, and relation in the k-th triple, respectively. and Respectively represent e k1 and r kr The matrices obtained after the 2D reshape operation in step (1-5-1) are respectively, * represents the convolution operation, f is the relu function, W is the parameter matrix of the fully connected layer, σ is the logistic sigmoid function, p(e k1 , r kr , e kt ) represents the prediction score of the k-th triple in the training set.

[0037] Preferably, the loss function L in step (1-6) is specifically:

[0038]

[0039] Where K represents the number of all triplets in the training set, t k Indicates whether there is a relationship between the head entity and the tail entity in the kth triple in the training set. If there is a relationship between the two, then t k is 1 if the value is true, otherwise it is 0.

[0040] Preferably, step (4) is specifically as follows: first, the frequency tf of the jth term in the i-th technical field is calculated. i，j , which is calculated using the following formula, where j∈[1, the total number of entries in all patent texts in the i-th technical field], i∈[1, K]:

[0041]

[0042] where n i，j represents the number of occurrences of the jth term in the patent text in the i-th technical field, ∑ K n K，j Represents the number of occurrences of the jth term in the patent texts in all K technical fields.

[0043] Then, calculate the inverse document frequency (idf) of the jth term in the i-th technical field j , which uses the following formula:

[0044]

[0045] Where |D| is the total number of patent texts in all K technical fields obtained from the patent database in step (4), t ij represents the jth term in the i-th technical field, d i represents the total number of patent texts in the i-th technical field, |{d i |t ij ∈d i}| represents the number of patent texts containing the jth term in the i-th technical field.

[0046] Then, calculate tf i，j ×idf j The product of (which is used to evaluate the importance of the j-th term in the i-th technical field to the patents in the i-th technical field) is taken as the importance of the j-th term in the i-th technical field;

[0047] Finally, from the importance of all entries in the i-th technical field, the entries corresponding to the largest Num importance are selected, and these Num entries are used as key technical words in the i-th technical field, where the value range of Num is between 1 and 20.

[0048] According to another aspect of the present invention, a system for generating a technology list based on a knowledge graph is provided, comprising:

[0049] The first module is used to obtain patent data, input the patent data into the pre-built knowledge graph and embed it into the neural network model to obtain the embedding result of the patent data, that is, the embedding vector of the patent data;

[0050] The second module is used to process the embedded vectors of the patent data obtained by the second module using the Canopy clustering algorithm to obtain K cluster center points;

[0051] The third module is used to use the K cluster center points obtained in the second module as initial center points and use the K-means algorithm to cluster the entity embedding vectors obtained in the first module to obtain K clustering results as K technical fields.

[0052] The fourth module is used to obtain all patent texts in each technical field from the patent database for each of the K technical fields obtained in the third module, and use the term frequency-inverse text frequency index TF-IDF algorithm to process all patent texts in each technical field to obtain multiple key technical words corresponding to the technical field.

[0053] The fifth module is used to obtain the corresponding technology research heat evaluation index and / or technology frontier evaluation index, as well as the value corresponding to the technology research heat evaluation index and / or technology frontier evaluation index for each of the K technology fields obtained in the third module, and then sort all K technology fields in descending order according to the values ​​corresponding to the technology research heat evaluation index and / or technology frontier evaluation index corresponding to all K technology fields, and finally generate a mapping relationship between the technology research heat evaluation index and / or technology frontier evaluation index, the sorted K technology fields, and the key technical terms corresponding to each technology field in the fourth module, thereby forming a technology list and outputting it.

[0054] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects compared with the prior art:

[0055] 1. The present invention adopts step (1-1), which uses a large amount of patent data to generate a technology list. Therefore, it can solve the defects of the existing methods, such as involving many experts, taking a long time, consuming a lot of energy, relying heavily on the knowledge and experience of experts, and lacking quantitative analysis and data support.

[0056] 2. Since the present invention adopts steps (1), (2) and (3), it dynamically divides the technical field by using the Canopy and K-means clustering methods, thereby being able to solve the defects existing in the existing technical field division methods, such as the rigidification of technical field division standards, the difficulty in discovering emerging technical fields, the scope and pertinence of the division standards sometimes being difficult to meet actual needs, and the inability to change with the development of technology.

[0057] 3. Since the present invention adopts step (4), it obtains technical keywords for each technical field through the word frequency-inverse text frequency index algorithm, and uses the technical keywords to describe the technical field. Therefore, it can solve the defects of the description method of the existing technical field that the new technical field generated by the technical integration of multiple technical fields cannot be well described, and the description of the technical field cannot change with the development of technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 It is a flow chart of the method for generating a technology list based on the knowledge graph of the present invention;

[0059] Figure 2 is a schematic diagram of entities and relationships in the knowledge graph data of the present invention;

[0060] Figure 3 It is a schematic diagram of the knowledge graph embedded in the neural network model of the present invention. DETAILED DESCRIPTION

[0061] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0062] The basic idea of ​​the present invention is to provide a method for generating a technology list based on a knowledge graph, which performs cluster analysis on patent data, combines international technology list data items with expert domain knowledge, and ultimately formulates a method for generating a technology list in the corresponding field under the guidance of a knowledge graph.

[0063] The present invention proposes a method and system for generating a technology list based on a knowledge graph. First, a patent knowledge graph is constructed for patent data. Then, patent entities are clustered. Then, the key technical terms of each field are obtained through the TF-IDF algorithm. Finally, a process method for formulating a technology list for the corresponding field is implemented in combination with evaluation indicators.

[0064] like Figure 1The present invention provides a method for generating a technology list based on a knowledge graph, comprising the following steps:

[0065] (1) Obtain patent data, input the patent data into a pre-built knowledge graph and embed it into a neural network model to obtain an embedding result of the patent data, i.e., an embedding vector of the patent data;

[0066] Specifically, the patent data in this step is obtained from a patent database such as Derwent.

[0067] The knowledge graph embedding neural network model in this step is the ConvE model (such as Figure 3 It is constructed by following the steps below:

[0068] (1-1) Obtaining patent datasets from patent databases;

[0069] Specifically, this step is to obtain a patent dataset from a patent database such as Derwent, and the patent dataset should contain as large a number of patents as possible (at least more than 100,000);

[0070] The advantage of this step is that it uses a large amount of patent data, thus getting rid of the heavy reliance on expert knowledge and experience, saving manpower and material resources, and providing quantitative analysis and data support.

[0071] (1-2) Generate multiple triples based on the patent dataset obtained in step (1-1), all triples constitute a knowledge graph dataset, and each knowledge graph data in the knowledge graph dataset is a triple (e m1 , r mr , e m2 ), where e m1 represents the embedding vector of the head entity in the mth triplet in the knowledge graph dataset, r mr Represents the embedding vector of the relationship in the mth triple in the knowledge graph dataset, e m2 Represents the embedding vector of the tail entity in the mth triple in the knowledge graph dataset, and m∈[1, the total number of triples in the knowledge graph dataset].

[0072] Specifically, the triples in this step are composed of head entity-relationship-tail entity, where there are five types of entities, including patent name, inventor, application unit, IPC classification number, and country, and five types of relationships, including application unit-patent name, patent name-IPC classification number, inventor-application unit, inventor-patent name, and application unit-country. Figure 2There are five types of triples generated in this step, namely (applicant-application-patent name), (patent name-belongs to-IPC classification number), (inventor-belongs to-applicant), (inventor-application-patent name), and (applicant-belongs to-country).

[0073] (1-3) Divide the knowledge graph data set obtained in step (1-2) into a training set, a validation set, and a test set. All entities in all knowledge graph data (i.e., triples) in the training set constitute a set E, and all relations in all knowledge graph data (i.e., triples) in the training set constitute a set R;

[0074] In this step, the knowledge graph training set is divided into the following proportions: training set 90%, validation set 5%, and test set 5%.

[0075] (1-4) Randomly initialize the entity set E and the relationship set R corresponding to the training set obtained in step (1-3) to obtain the embedding vector of each entity and each relationship;

[0076] Specifically, the size of the embedding vector is d, the value range of d is 1 to 1024, preferably 200, and each embedding vector satisfies a normal distribution with a mean of 0;

[0077] (1-5) Input the embedding vectors of all entities and relationships in the training set obtained in step (1-4) into the knowledge graph embedding neural network model ConvE to obtain multiple prediction scores;

[0078] Specifically, this step includes the following sub-steps:

[0079] (1-5-1) For each triple in the training set (e k1 , r kr , e k2 ), the 2D reshape operation is used to transform the triple (e k1 , r kr , e k2 ) is the embedding vector e of the head entity k1 and the embedding vector r of the relation kr Convert into a matrix and concatenate to get the triple (e k1 , r kr , e k2 ) where k∈[1, the total number of triplets in the training set].

[0080] (1-5-2) For each triple in the training set (e k1 , r kr , e k2), the triplet (e k1 , r kr , e k2 )The corresponding concatenated matrix is ​​input into the two-dimensional convolutional layer with filter ω in the ConvE model to obtain the tensor corresponding to the triplet.

[0081] (1-5-3) For each triple in the training set (e k1 , r kr , e k2 ), the tensor corresponding to the triplet is flattened into a one-dimensional vector and then input into the fully connected layer with parameter matrix W in the ConvE model to obtain the triplet (e k1 , r kr , e kt ) corresponds to the output vector.

[0082] (1-5-4) For each triple in the training set, add the vector corresponding to the triple obtained in step (1-5-3) to the embedding vector e of the tail entity in the triple kt Perform the inner product operation and use the logistic sigmoid function to process the inner product calculation result to obtain the output result as the triple (e1, r r ,e t )’s prediction score;

[0083] Specifically, this step uses the following formula:

[0084]

[0085] where e k1 、e k2 and r kr Represent the embedding vectors of the head entity, tail entity, and relation in the k-th triple, respectively. and Respectively represent e k1 and r kr The matrices obtained after the 2D reshape operation in step (1-5-1) are respectively, * represents the convolution operation, f is the relu function, W is the parameter matrix of the fully connected layer, σ is the logistic sigmoid function, p(e k1 , r kr , e k2 ) represents the prediction score of the k-th triple in the training set.

[0086] (1-6) Input all the prediction scores obtained in step (1-5) into the loss function to minimize the loss function L, thereby obtaining a preliminarily trained knowledge graph embedded in the neural network model;

[0087] The loss function L is specifically:

[0088]

[0089] Where K represents the number of all triplets in the training set, t k Indicates whether there is a relationship between the head entity and the tail entity in the kth triple in the training set (for example, if the head entity is the application unit and the tail entity is the patent name, there is a relationship between them; if the head entity is the application unit and the tail entity is the IPC classification number, there is no relationship between them). If there is a relationship between the two, then t k is 1 if the value is true, otherwise it is 0.

[0090] (1-7) Verifying the knowledge graph embedding neural network model preliminarily trained in step (1-6) according to the verification set obtained in step (1-3) until the loss function L is optimal, thereby obtaining a trained knowledge graph embedding neural network model;

[0091] (2) Using the Canopy clustering algorithm to process the embedding vector of the patent data obtained in step (1) to obtain K cluster centers; (3) Using the K cluster centers obtained in step (2) as the initial centers, using the K-means algorithm to cluster the entity embedding vector obtained in step (1) to obtain K clustering results as K technical fields.

[0092] Specifically, in this step, the patent data is classified by clustering to obtain K types of patent data. Similar patent data are in the same category, and the patent data in the same category are regarded as patent data belonging to the same technical field. In this way, the division of technical fields is obtained, that is, K technical fields;

[0093] The advantages of the above steps (1) to (3) are that they can solve the problems existing in the existing technology field division methods, such as the rigidity of division standards, the difficulty in discovering emerging technology fields, the scope and specificity of the division standards sometimes being unable to meet actual needs, and the inability to change with technological development.

[0094] (4) For each of the K technical fields obtained in step (3), all patent texts in each technical field are obtained from a patent database (e.g., the Derwent patent database), and all patent texts in each technical field are processed using a term frequency-inverse document frequency (TF-IDF) algorithm to obtain multiple key technical terms corresponding to the technical field.

[0095] Specifically, in this step, the frequency tf of the jth term in the i-th technical field is calculated first. i，j (where j∈[1, the total number of entries in all patent texts in the i-th technical field], i∈[1, K]), which is calculated using the following formula:

[0096]

[0097] where n i，j represents the number of occurrences of the jth term in the patent text in the i-th technical field, ∑ K n K，j Represents the number of occurrences of the jth term in the patent texts in all K technical fields.

[0098] Then, calculate the inverse document frequency (idf) of the jth term in the i-th technical field j , which uses the following formula:

[0099]

[0100] Where |D| is the total number of patent texts in all K technical fields obtained from the patent database in step (4), t ij represents the jth term in the i-th technical field, d i represents the total number of patent texts in the i-th technical field, |{d i |t ij ∈d i}| represents the number of patent texts containing the jth term in the i-th technical field.

[0101] Then, calculate tf i，j ×idf j The product of (which is used to evaluate the importance of the j-th term in the i-th technical field to the patents in the i-th technical field) is taken as the importance of the j-th term in the i-th technical field;

[0102] Finally, from the importance of all terms in the i-th technical field, select the terms corresponding to the largest Num importances, and use these Num terms as key technical terms in the i-th technical field, where the value range of Num is between 1 and 20, preferably 10.

[0103] The advantages of this step are that it can solve the problems that the existing description methods of technical fields cannot well describe new technical fields generated by the fusion of multiple technical fields, and the description of technical fields cannot change with the development of technology.

[0104] (5) For each of the K technical fields obtained in step (3), obtain its corresponding technology research heat evaluation index and / or technology frontier evaluation index, as well as the value corresponding to the technology research heat evaluation index and / or technology frontier evaluation index, and then sort all K technical fields in descending order according to the values ​​corresponding to the technology research heat evaluation index and / or technology frontier evaluation index corresponding to all K technical fields, and finally generate a mapping relationship between the technology research heat evaluation index and / or technology frontier evaluation index, the sorted K technical fields, and the key technical terms corresponding to each technical field in step (4), thereby forming a technology list and outputting it.

[0105] Specifically, as shown in Table 1 below, the technical research heat evaluation indicators of a certain technical field include the total number of authorized patents in the technical field, the total number of US patent authorizations, the Current Influence Index (CII), technical strength (TS), technical influence index (TII), forward citations, and scientific relevance.

[0106] Table 1

[0107]

[0108] As shown in Table 2 below, the evaluation indicators of technological frontier in a certain technical field include the technology maturity coefficient, technology aging coefficient, invention patent authorization rate, technology growth rate, and burst value in the technical field.

[0109] Table 2

[0110]

[0111] For example, if the desired technology list is related to the total number of U.S. patent authorizations, then in this step, first, the total number of U.S. patent authorizations corresponding to each of the K technical fields is obtained respectively, and then the K technical fields are sorted in descending order of the total number of U.S. patent authorizations (for example, the order is K-1th technical field, 2nd technical field, 4th technical field, 1st technical field, Kth technical field, ..., 3rd technical field), and then a technology list is established, which includes three items: the indicator name (i.e., the total number of U.S. patent authorizations), the sorted K technical fields, and the key technical words corresponding to each technical field, as shown in Table 3 below (assuming K=4, Num=2):

[0112] Table 3

[0113]

[0114]

[0115] It will be easily understood by those skilled in the art that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for generating a technology list based on a knowledge graph, characterized in that: The following steps are involved: (1) Obtain patent data, input the patent data into a pre-built knowledge graph and embed it into a neural network model to obtain an embedding result of the patent data, i.e., an embedding vector of the patent data; (2) Using the Canopy clustering algorithm to process the embedded vectors of the patent data obtained in step (1) to obtain K cluster center points; (3) Taking the K cluster center points obtained in step (2) as the initial center points, the entity embedding vectors obtained in step (1) are clustered using the K-means algorithm to obtain K clustering results as K technical fields; (4) For each of the K technical fields obtained in step (3), all patent texts in each technical field are obtained from the patent database, and all patent texts in each technical field are processed using a term frequency-inverse text frequency index TF-IDF algorithm to obtain multiple key technical terms corresponding to the technical field; (5) For each of the K technical fields obtained in step (3), obtain its corresponding technology research heat evaluation index and / or technology frontier evaluation index, as well as the value corresponding to the technology research heat evaluation index and / or technology frontier evaluation index, and then sort all K technical fields in descending order according to the values ​​corresponding to the technology research heat evaluation index and / or technology frontier evaluation index corresponding to all K technical fields, and finally generate a mapping relationship between the technology research heat evaluation index and / or technology frontier evaluation index, the sorted K technical fields, and the key technical terms corresponding to each technical field in step (4), thereby forming a technology list and outputting it.

2. The method for generating a technology list based on a knowledge graph according to claim 1, characterized in that: The knowledge graph embedding neural network model is a ConvE model and is constructed through the following steps: (1-1) Obtaining patent datasets from patent databases; (1-2) Generate multiple triples based on the patent dataset obtained in step (1-1), all triples constitute a knowledge graph dataset, and each knowledge graph data in the knowledge graph dataset is a triple (e m1 ,r mr ,e m2 ), where e m1 represents the embedding vector of the head entity in the mth triplet in the knowledge graph dataset, r mr Represents the embedding vector of the relationship in the mth triple in the knowledge graph dataset, e m2 Represents the embedding vector of the tail entity in the mth triple in the knowledge graph dataset, and m∈[1, the total number of triples in the knowledge graph dataset]; (1-3) Divide the knowledge graph data set obtained in step (1-2) into a training set, a validation set, and a test set. All entities in all knowledge graph data in the training set constitute a set E, and all relations in all knowledge graph data in the training set constitute a set R; (1-4) Randomly initialize the entity set E and the relationship set R corresponding to the training set obtained in step (1-3) to obtain the embedding vector of each entity and each relationship; (1-5) Input the embedding vectors of all entities and relationships in the training set obtained in step (1-4) into the knowledge graph embedding neural network model ConvE to obtain multiple prediction scores; (1-6) Input all the prediction scores obtained in step (1-5) into the loss function to minimize the loss function L, thereby obtaining a preliminarily trained knowledge graph embedded in the neural network model; (1-7) The knowledge graph embedding neural network model preliminarily trained in step (1-6) is verified according to the verification set obtained in step (1-3) until the loss function L is optimal, thereby obtaining a trained knowledge graph embedding neural network model.

3. The method for generating a technology list based on a knowledge graph according to claim 1 or 2, characterized in that: The triple in step (1-2) is composed of head entity-relation-tail entity; Entities include patent title, inventor, applicant, IPC classification number, and country; The relationships include applicant-patent name, patent name-IPC classification number, inventor-applicant, inventor-patent name, and applicant-country; The triples generated in step (1-2) include (applicant unit-application-patent name), (patent name-belongs to-IPC classification number), (inventor-belongs to-applicant unit), (inventor-application-patent name), and (applicant unit-belongs to-country).

4. The method for generating a technology list based on a knowledge graph according to any one of claims 1 to 3, characterized in that: The value range of the embedding vector in step (1-4) is 1 to 1024; Each embedding vector satisfies a normal distribution with a mean of 0.

5. The method for generating a technology list based on a knowledge graph according to claim 1, characterized in that: Step (1-5) includes the following sub-steps: (1-5-1) For each triple in the training set (e k1 ,r kr ,e k2 ), the 2D reshape operation is used to transform the triple (e k1 ,r kr ,e k2 ) is the embedding vector e of the head entity k1 and the embedding vector r of the relation kr Convert into a matrix and concatenate to get the triple (e k1 ,r kr ,e k2 ) corresponding to the concatenation matrix, where k∈[1, the total number of triplets in the training set]; (1-5-2) For each triple in the training set (e k1 ,r kr ,e k2 ), the triplet (e k1 ,r kr ,e k2 ) is input into the two-dimensional convolutional layer with filter ω in the ConvE model to obtain the tensor corresponding to the triplet; (1-5-3) For each triple in the training set (e k1 ,r kr ,e k2 ), the tensor corresponding to the triplet is flattened into a one-dimensional vector and then input into the fully connected layer with parameter matrix W in the ConvE model to obtain the triplet (e k1 ,r kr ,e k2 ) corresponding to the output vector; (1-5-4) For each triple in the training set, the vector corresponding to the triple obtained in step (1-5-3) is compared with the embedding vector e of the tail entity in the triple k2 Perform the inner product operation and use the logistic sigmoid function to process the inner product calculation result to obtain the output result as the triple (e1, r r ,e2)’s prediction score.

6. The method for generating a technology list based on a knowledge graph according to claim 5, characterized in that: The prediction score in step (1-5-4) is calculated using the following formula: where e k1 、e k2 and r kr Represent the embedding vectors of the head entity, tail entity, and relation in the k-th triple, respectively. and Respectively represent e k1 and r kr The matrices obtained after the 2D reshape operation in step (1-5-1) are respectively, * represents the convolution operation, f is the relu function, W is the parameter matrix of the fully connected layer, σ is the logistic sigmoid function, p(e k1 ,r kr ,e k2 ) represents the prediction score of the k-th triple in the training set.

7. The method for generating a technology list based on a knowledge graph according to claim 5, characterized in that: The loss function L in steps (1-6) is specifically: Where K represents the number of all triplets in the training set, t k Indicates whether there is a relationship between the head entity and the tail entity in the kth triple in the training set. If there is a relationship between the two, then t k is 1 if the value is true, otherwise it is 0.

8. The method for generating a technology list based on a knowledge graph according to claim 1, characterized in that: Step (4) is as follows: first, calculate the frequency tf of the jth term in the i-th technical field. i,j , which is calculated using the following formula, where j∈[1, the total number of entries in all patent texts in the i-th technical field], i∈[1, K]: where n i,j represents the number of occurrences of the jth term in the patent text in the i-th technical field, ∑ K n K,j represents the number of occurrences of the jth term in the patent texts in all K technical fields; Then, calculate the inverse document frequency (idf) of the jth term in the i-th technical field j , which uses the following formula: Where |D| is the total number of patent texts in all K technical fields obtained from the patent database in step (4), t ij represents the jth term in the i-th technical field, d i represents the total number of patent texts in the i-th technical field, |{d i |t ij ∈d i }| represents the number of patent texts containing the jth term in the i-th technical field; Then, calculate tf i,j ×idf j The product of is taken as the importance of the jth term in the i-th technical field; Finally, from the importance of all entries in the i-th technical field, the entries corresponding to the largest Num importance are selected, and these Num entries are used as key technical words in the i-th technical field, where the value range of Num is between 1 and 20.

9. A technology list generation system based on knowledge graph, characterized in that: include: The first module is used to obtain patent data, input the patent data into the pre-built knowledge graph and embed it into the neural network model to obtain the embedding result of the patent data, that is, the embedding vector of the patent data; The second module is used to process the embedded vectors of the patent data obtained by the second module using the Canopy clustering algorithm to obtain K cluster center points; The third module is used to use the K cluster center points obtained in the second module as initial center points and use the K-means algorithm to perform clustering processing on the entity embedding vectors obtained in the first module to obtain K clustering results as K technical fields; The fourth module is used for obtaining all patent texts in each technical field from the patent database for each of the K technical fields obtained in the third module, and processing all patent texts in each technical field using the term frequency-inverse text frequency index TF-IDF algorithm to obtain multiple key technical words corresponding to the technical field; The fifth module is used to obtain the corresponding technology research heat evaluation index and / or technology frontier evaluation index, as well as the value corresponding to the technology research heat evaluation index and / or technology frontier evaluation index for each of the K technology fields obtained in the third module, and then sort all K technology fields in descending order according to the values ​​corresponding to the technology research heat evaluation index and / or technology frontier evaluation index corresponding to all K technology fields, and finally generate a mapping relationship between the technology research heat evaluation index and / or technology frontier evaluation index, the sorted K technology fields, and the key technical terms corresponding to each technology field in the fourth module, thereby forming a technology list and outputting it.

Images