Core technology identification method based on dual technology relevance

Abstract

The invention discloses a core technology identification method based on dual technology relevance, which comprises the following steps of: firstly, collecting related technical documents, and preprocessing text contents in the documents; secondly, word embedding corresponding to lexical items contained in the technical document is obtained through Word2Vec model training; then, taking the obtained technical documents subjected to word embedding and preprocessing as input, and extracting themes representing different technologies through an ETM-based theme modeling process; secondly, measuring the proximity and the combination association degree between the topics by using a cosine similarity and an assertion string standardization method; next, the topics and the proximity and combination association between the topics are respectively used as nodes and node links in the network to construct a dual network # imgabs0 #; a multi-centrality analysis method based on a plurality of centrality indexes and factor analysis is used to identify a dual core technology, a core combination technology, a core proximity technology and an edge technology in a dual network # imgabs1 #.

Description

Technical Field

[0001] The invention relates to a core technology identification method based on dual technology relevance. Background Art

[0002] Technological relevance shapes the technological diversification model of innovation subjects and affects the development direction of technological innovation. Clarifying the relationship between technologies helps to clarify the relative importance of different technologies and the correlation structure in the field of technology. Exploring the relevance between different technologies can help innovation subjects fully grasp technological dynamics and assist in making implementation decisions for technological innovation.

[0003] Existing studies have ignored the multiple types of associations of technological associations and only considered a single type of relationship between technologies. However, the concept of technological associations is not a single dimension, but contains multiple types of associations. Different types of associations have different association strengths between the same set of technologies and show complex association patterns, which contain richer technological information than when only a single type of association is considered. Considering multiple types of technological associations can help reveal the true relationship structure between multiple technological fields in a technological system, so that innovation subjects can make correct technological innovation decisions.

[0004] Current research methods all rely on single-layer networks to study a single correlation between technologies. Limited by the model capabilities of single-layer networks, these methods can only study one type of correlation and cannot explore multiple types of correlations. When using such methods to study multiple correlations, multiple correlation classes can only be aggregated into one, but this will lose the correlation information contained in different types of correlations. Therefore, a method is needed to simultaneously study and reveal multiple correlations between technologies.

[0005] The measurement of the degree of association between technologies is the basis for studying the relevance of technologies, and it determines the correlation structure between technologies. Most of the existing methods for measuring the degree of association between technologies do not use text-based methods. Although some methods use keyword-based techniques to try to use text data to achieve more accurate association measurement, keyword-based methods can only capture discrete relationships between keywords, while ignoring semantic connections. In addition, such methods cannot handle the frequent appearance of technical terms, synonyms, and polysemous words in patent texts. Therefore, a text-based method that can take into account text semantics is needed to achieve more accurate association measurement.

[0006] In addition, most of the existing core technology identification methods only consider the relative importance of technology from a single dimension and fail to measure the importance of technology from a complete perspective. This may lead to the identification of one-sided core technologies and ignore other potential core technologies. Therefore, there is an urgent need for a method that can measure the relative importance of technologies from multiple dimensions and identify core technologies in a more comprehensive way. Summary of the invention

[0007] The purpose of the present invention is to provide a core technology identification method based on dual technology relevance to solve the technical problems raised by the background technology, comprising the following steps:

[0008] Extract technical keywords based on technical documents, construct a technical search formula, and obtain D original technical documents from the technical database based on the technical search formula;

[0009] Performing sentence segmentation and a first text preprocessing on the text content in each of the D original technical documents to obtain preprocessed technical document sentences, and training a word embedding model based on the preprocessed technical document sentences to obtain word embeddings corresponding to the terms contained in the D original technical documents;

[0010] Performing a second text preprocessing based on the D original technical documents to obtain technical documents recorded in the form of word bags;

[0011] Based on the word embedding and the technical documents in the bag-of-words form, topic modeling is performed through the ETM model to obtain three results, including a topic-term probability distribution matrix β, a document-topic probability distribution matrix θ, and a topic embedding matrix α;

[0012] Based on the topic embedding matrix α, the cosine similarity calculation formula is used to measure the proximity between two topic embeddings, and the proximity association matrix P that records the proximity association strength between topics is obtained. Based on the document-topic probability distribution matrix θ, the combined association matrix C that records the combined association strength between topics is calculated by the association strength standardization method. Based on the proximity association matrix P and the combined association matrix C, a dual network with two network levels is constructed. And for dual network Visualize; identify dual networks using multiple centrality analysis methods based on multiple centrality indicators and factor analysis Dual-core technology, core combination technology, core adjacent technology and edge technology in the industry;

[0013] Dual Network It can be represented by the hyperadjacency matrix A:

[0014]

[0015] in,

[0016]

[0017] The dual core technologies refer to topics that show high centrality in network positions both in the proximity network layer and in the combined network layer;

[0018] The core combination technology refers to a topic with low centrality in the proximity network layer and high centrality in the combination network layer;

[0019] The edge technologies refer to topics that show low centrality in both the proximity network layer and the combination network layer;

[0020] The core proximity technique refers to topics that have high centrality in the proximity network layer and low centrality in the combination network layer.

[0021] Furthermore, the word embedding model is a Word2Vec model or a glove model.

[0022] Furthermore, the specific steps of sentence segmentation and the first text preprocessing include:

[0023] The text content of each original technical document is segmented into technical document sentences with sentences as units through the Python toolkit NLTK.

[0024] The python toolkit spaCy is used to perform the first text preprocessing on the technical document after sentence segmentation to obtain the preprocessed technical document sentences.

[0025] Furthermore, the first text preprocessing process: first, delete numbers, punctuation marks and words with a character length less than 3 in the text; then, restore the remaining words to lowercase; finally, delete sentences whose sentence length is less than the "window size" parameter value set by the subsequent Word2Vec model to obtain preprocessed technical document sentences.

[0026] Furthermore, the second text preprocessing uses the python toolkit spaCy, and the processing process includes deleting numbers, punctuation marks, stop words, terms with a character length of less than 3, and non-noun terms in the original technical document, restoring the remaining terms and converting them to lowercase, deleting terms with a document frequency of less than 3 in the technical document collection, and finally obtaining the preprocessed technical document recorded in the form of a bag of words with a dictionary size of V.

[0027] Furthermore, based on the word embedding and the technical documents in the form of bag-of-words, topic modeling is performed through the ETM model to obtain three results, including the topic-term probability distribution matrix β, the document-topic probability distribution matrix θ and the topic embedding matrix α.

[0028] The specific steps are as follows:

[0029] Based on the word embedding and the technical documents recorded in the bag-of-words form as input data for training the ETM model, the optimal setting value of the parameter of the number of topics K in the ETM model is determined through repeated experiments;

[0030] The technical documents recorded in the word embedding and bag-of-words form are used as model inputs, and the topic modeling process is performed based on the ETM model parameter value topic number K to obtain three output results in the form of matrices:

[0031] Topic-term probability distribution matrix β:

[0032]

[0033] Row vector β k· is the probability distribution of topic k on all terms, where k is the kth topic among K topics;

[0034] Document-topic probability distribution matrix θ:

[0035]

[0036] Row vector θ d· is the probability distribution of document d on all topics, where d is the d-th document in D topics;

[0037] Topic embedding matrix α:

[0038]

[0039] The column vector α·k is the topic embedding corresponding to topic k, where k is the kth topic among K topics.

[0040] Furthermore, the cosine similarity calculation formula is used to measure the proximity between two topic embeddings. The specific steps to obtain the proximity association matrix P that records the proximity association strength between topics are as follows:

[0041] Use the cosine similarity calculation formula to calculate the element values ​​in the proximity association matrix P;

[0042]

[0043] P is a symmetric square matrix, and the calculation formula of element value is:

[0044]

[0045] Among them, α i is the topic embedding corresponding to topic i, α j is the topic embedding corresponding to topic j, |α i || is the L-2 norm of the topic embedding corresponding to topic i, ||α j || is the L-2 norm of the topic embedding corresponding to topic j, p ij and p ji is the proximity correlation degree between topic i and topic j;

[0046] The range of element values ​​in the proximity association matrix P is scaled from [-1, 1] to [0, 1] by standardization. The standardization calculation formula is:

[0047]

[0048] p′ ij is the normalized matrix element value;

[0049] The diagonal element values ​​of the proximity association matrix P are returned to 0, that is:

[0050] p′ ii =0

[0051] Finally, the processed proximity correlation matrix is ​​obtained.

[0052] Furthermore, the specific steps of calculating the combined association matrix C that records the combined association strengths between topics based on the document-topic probability distribution matrix θ by the association strength standardization method are as follows:

[0053] Set a threshold η and set the element values ​​less than η in the document-topic probability distribution matrix θ to 0, thereby filtering rare events. The formula is expressed as:

[0054]

[0055] The combined association matrix C of the combined association strength between the recorded topics is calculated using the association strength standardization method:

[0056]

[0057] C is a symmetric square matrix, and the calculation formula of element value is:

[0058]

[0059] θ idis the element value in the document-topic probability distribution matrix θ, that is, the probability that topic i appears in document d; θ jd is the element value in the document-topic probability distribution matrix θ, that is, the probability that topic j appears in a document d; ∑ d θ id θ jd is the probability that topic i and topic j appear together in the document collection, ∑ d θ id is the total probability of topic i appearing in the document collection; ∑ d θ jd is the total probability of topic j appearing in the document collection;

[0060] The diagonal element values ​​of the combined correlation matrix C are returned to zero, and finally the processed combined correlation matrix C is obtained. The return-to-zero processing formula is:

[0061] c ii =0.

[0062] Furthermore, the steps of using a multiple centrality analysis method based on multiple centrality indicators and factor analysis to identify dual core technologies, core combination technologies, core adjacent technologies and edge technologies in the dual network M include:

[0063] Calculate the strength of each topic i in the proximity network layer among the K topics The calculation formula is: Calculate the strength of each topic i in the combined network layer among the K topics The calculation formula is: The dual network is calculated using the python toolkit NetworkX The feature vector centrality of each topic i in the proximity network layer and the combination network layer in the K topics in and Current Betweenness Centrality and Current proximity to centrality and

[0064] Among them, the eigenvector centrality of each topic i in the p network layer corresponding to the proximity association matrix P is obtained through the eigenvector_centrality function in the python toolkit NetworkX The eigenvector_centrality function in the Python toolkit NetworkX is used to obtain the eigenvector centrality of each topic i in the c network layer corresponding to the combined correlation matrix C. The current betweenness centrality of each topic i in the p network layer corresponding to the proximity association matrix P is obtained through the current_flow_betweenness_centrality function in the python toolkit NetworkX The current betweenness centrality of each topic i in the c network layer corresponding to the combined correlation matrix C is obtained through the current_flow_betweenness_centrality function in the python toolkit NetworkX The current closeness centrality of each topic i in the p network layer corresponding to the proximity association matrix P is obtained through the current_flow_closeness_centrality function in the Python toolkit NetworkX The current closeness centrality of each topic i in the combined network layer corresponding to the combined correlation matrix C is obtained through the current_flow_closeness_centrality function in the Python toolkit NetworkX

[0065] For each topic i among the K topics As input data, factor analysis was performed using the factor function of stata software to obtain the centrality variables that can explain the four layers of the proximity network. Commonality factor f p And each topic i among the K topics has a factor f p The corresponding factor score f p (i) and the centrality variables that can explain the four combinations of network layers Commonality factor f c And each topic i among the K topics has a factor f c The corresponding factor score f c (i), each subject i in the dual network The centrality in is represented as a vector

[0066] Set the factor score f for the partition topic i p (i) High and low thresholds γ p , set the factor score f of the partition topic i c (i) High and low thresholds γ c , which is used to judge the centrality of the network position of subject i in the proximity network layer and the combination network layer respectively;

[0067] The factor score f p (i) As the x-axis, the factor score fc (i) As the y-axis, a two-dimensional coordinate system is constructed, and based on the vector Construct a two-dimensional space based on two thresholds γ p and γ c Divide the two-dimensional space into four spatial regions to identify the dual network Dual-core technology, core combination technology, core adjacent technology and edge technology.

[0068] Furthermore, the threshold γ p Set to the median or third quantile of the factor scores of K topics in the proximity network layer; the threshold γ c Set to the median or third quantile of the factor scores of the K topics in the combined network layer.

[0069] The beneficial effect of the present invention is that the method can embed two types of technical associations, proximity and combination, in a network system, visualize the association structure between technologies, and analyze various association patterns between technologies. First, collect relevant technical documents and preprocess the text content in the documents; second, obtain the word embeddings corresponding to the terms contained in the technical documents through Word2Vec model training; then, use the obtained word embeddings and the preprocessed technical documents as input, and extract topics representing different technologies through ETM-based topic modeling; then, use the cosine similarity and association strength standardization methods to measure the proximity and combination association between topics; next, use the proximity and combination associations between topics as nodes and node links in the network to construct a dual network. Finally, the multi-centrality analysis method based on multiple centrality indicators and factor analysis identified dual networks. Dual-core technology, core combination technology, core adjacent technology and edge technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0071] Figure 1 It is a topic consistency display diagram of the ETM model corresponding to different number of topics K;

[0072] Figure 2 This is a graph showing the topic embedding of topic 5 and the word embeddings of the 15 terms closest to its cosine;

[0073] Figure 3 It is a dual network display diagram embedded with dual technology associations;

[0074] Figure 4 Identify dual networks A two-dimensional spatial display of dual core technology, core combination technology, core adjacent technology and edge technology.

[0075] It should be noted that the drawings are not necessarily drawn to scale, but are merely shown in a schematic manner that does not affect the reader's understanding. DETAILED DESCRIPTION

[0076] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0077] The present invention provides a core technology identification method based on dual technology relevance, comprising the following steps:

[0078] S1: Extract technical keywords based on technical documents and construct technical search formulas;

[0079] For example, the technical documents are multiple pieces of information based on the field of 3D printing technology. Keywords describing the technical field are extracted from these documents, and a technical search formula suitable for the technical database is constructed. For example, a specific search formula suitable for the incoPat patent database is as follows:

[0080] “TIAB=(3D printing) OR TIAB=(three-dimensional printing) OR TIAB=(3DP) OR TIAB=(additive manufacturing) ORTIAB=(rapid prototyping) OR TIAB=(three-dimensional manufacturing)”

[0081] S2: Obtain D original technical documents from the technical database based on the technical search formula;

[0082] For example, the technical database is a patent database (incoPat patent database or DerwentInnovations Index patent database). For example, based on the technical search formula for 3D printing technology constructed in Example S1, 67,231 original technical documents are searched from the incoPat patent database. Preferably, the content of the original technical document is the title text and abstract text of the patent specification on which the subsequent method analysis is based.

[0083] S3: performing sentence segmentation and first text preprocessing on the text contents in the D original technical documents to obtain preprocessed technical document sentences;

[0084] Specifically, the specific steps of sentence segmentation and the first text preprocessing include:

[0085] S31: using the Python toolkit NLTK, the text content of each original technical document is segmented to obtain technical document sentences with sentences as units;

[0086] For example, the text contents of the 67,231 original technical documents obtained in Example S2 are segmented to obtain 590,634 technical document sentences.

[0087] S32: Use the python toolkit spaCy to perform the first text preprocessing on the technical document sentences after sentence segmentation to obtain the preprocessed technical document sentences.

[0088] Specifically, the first text preprocessing process: first, delete numbers, punctuation marks and words with a character length less than 3 in the text; then, perform part-of-speech restoration on the remaining words and convert them into lowercase; finally, delete sentences whose sentence length (the number of words contained in the sentence) is less than the "window size" parameter value set by the subsequent Word2Vec model to obtain preprocessed technical document sentences;

[0089] For example, the 590,634 technical document sentences obtained in Example S31 were subjected to text preprocessing to obtain 561,733 preprocessed technical document sentences.

[0090] S4: Based on the preprocessed technical document sentences obtained in S3, the Python toolkit Gensim is used to train the Word2Vec model to obtain word embeddings corresponding to D terms contained in the original technical document; the word embedding results will be used as one of the input data for subsequent ETM model training;

[0091] For example, when training the Word2Vec model, the parameters are set as follows: the model architecture is Skip-Gram, the negative sampling value is 15, the word embedding dimension is 300, the window size is 5, the number of iterations is 20, the minimum word frequency is 1, and the other parameters not mentioned use the default parameter values ​​of Gensim's Word2Vec model API. Based on the 561,733 technical document sentences obtained in the S32 example, the 300-dimensional word embedding corresponding to 19,709 terms is obtained through the Word2Vec model training with the above parameter settings.

[0092] Preferably, in the present invention, the word embedding model is a Word2Vec model. Of course, in addition to the above models, it can also be other word embedding models such as Glove.

[0093] First, the Word2Vec model is simple and efficient, and is suitable for limited resource scenarios. Because the Word2Vec model uses a shallow neural network to generate word embeddings, it is simpler and more efficient than complex models such as BERT, and it still performs well in situations where computing resources are limited, which allows researchers and practitioners to use it without a lot of hardware support.

[0094] Secondly, the Word2Vec model is fast to train. The Word2vec model uses some acceleration methods, such as NegativeSampling and Hierarchical Softmax, which makes it faster during training.

[0095] Finally, the Word2Vec model supports incremental training, which means that the Word2Vec model can be updated with new data without training from scratch, which is a very useful feature for applications that need to constantly introduce new information.

[0096] S5: Perform a second text preprocessing based on the D original technical documents described in S2 to obtain technical documents recorded in the form of word bags; the technical documents recorded in the form of word bags will be used as one of the input data for subsequent ETM model training;

[0097] Specifically, the second text preprocessing uses the python toolkit spaCy, and the processing process includes deleting numbers, punctuation marks, stop words, terms with a character length of less than 3, and terms with non-noun parts of speech in the original technical documents. The stop words refer to terms that appear frequently in the language and contribute less to the actual semantics of the text, such as articles, prepositions, etc. Finally, the remaining terms are restored to their parts of speech and converted to lowercase, and terms with a document frequency of less than 3 in the technical document collection are deleted. Finally, the preprocessed technical documents recorded in the form of a bag of words with a dictionary size of V are obtained. The dictionary refers to a collection of all unique (i.e., unique, non-repetitive) terms extracted from the technical documents.

[0098] For example, the 67,231 original technical documents obtained in Example S2 are subjected to a second text preprocessing to obtain 67,135 technical documents recorded in the form of bag-of-words with a dictionary size of 9042.

[0099] In the second text preprocessing process, choosing to retain only noun parts of speech can improve the training efficiency of the topic model and improve the semantic consistency and interpretability of the topic results. Technical-related components and terminology descriptions such as methods and tools usually appear in the form of nouns, so words of other parts of speech will become noise during the topic model training process, thereby affecting the model training efficiency and the accuracy of the results, and will not play a positive role in the interpretation of the meaning of the topic results. Therefore, in the second text preprocessing process, words other than noun parts of speech were deleted to solve these problems, which played a positive role in promoting the topic model training process.

[0100] S6: Based on the word embedding obtained in S4 and the technical documents in the bag-of-words form obtained in S5, topic modeling is performed through the ETM model to obtain three results, including the topic-term probability distribution matrix β, the document-topic probability distribution matrix θ and the topic embedding matrix α;

[0101] First, the ETM used in the present invention has higher operating efficiency and computational efficiency. For example, traditional topic models such as LSA (Latent Semantic Allocation) and LDA (Latent Dirichlet allocation) have low computational efficiency and operating efficiency. Traditional topic models have parameter reasoning processes specific to different models, and the complexity of the parameter reasoning process increases with the complexity of the model itself. In addition, traditional topic models consume a lot of time when facing large document sets. The ETM model used in the steps of this method is a topic model based on a neural network, so the model parameters can be inferred through an automatic gradient back-propagation process, and parallel computing is allowed, thereby greatly improving the computational efficiency of the topic modeling process.

[0102] Secondly, the ETM model used has better model interpretability. The interpretability of traditional topic models is poor, and the training process only relies on a single input data of a document set represented in the form of a bag of words. Since the document set represented in the form of a bag of words can only record the discrete relationship between terms in the text content, it cannot capture the semantic relationship reflected by the context in which the terms are located, so it will cause the loss of semantic information. The training process of the ETM model used in the method steps of the present invention not only uses a document set represented in the form of a bag of words, but also utilizes word embedding containing semantic information in the document set. Word embedding can encode the semantic and grammatical information of a term into its vector (i.e., embedding) representation. Word embedding, as the second input data of ETM, makes the ETM model more interpretable.

[0103] Furthermore, the ETM model used is robust to stop words and rare words. When faced with stop words and rare words, the interpretability of the model results of traditional topic models will decrease. Because stop words and rare words appear in the content of topic results, they are not helpful for the interpretation of topic results. However, when faced with meaningless and frequently occurring stop words and rare words, the ETM model used in this method will generate separate topics for such terms, distinguishing meaningful terms from meaningless stop words and rare words, thereby improving the interpretability of the model results.

[0104] Finally, thanks to the input word embedding data, the ETM model used has unique outputs that other topic models do not have——

[0105] Topic embedding. Traditional topic models have only two output results, namely the topic-term probability distribution matrix β and the document-topic probability distribution matrix θ. ETM has a unique output, namely the topic embedding matrix α. Topic embedding can express the semantic content and rich contextual semantic information contained in the topic through vectors (i.e. embedding), which provides a basis for more accurate measurement of the degree of association between topics.

[0106] The process of obtaining three result matrices through topic modeling using the ETM model includes the following steps:

[0107] S61: The word embedding obtained in S4 and the technical documents recorded in the form of bag-of-words obtained in S5 are used as input data for training the ETM model. The optimal setting value of the parameter K of the number of topics in the ETM model is determined through repeated experiments.

[0108] For example, in the ETM model, other parameters except the number of topics K remain unchanged, and different numbers of topics K are set to repeat the experiment, and the topic coherence (TC) index of each experimental model result is calculated. The higher the topic coherence value, the stronger the interpretability of the topic model. The topic coherence (TC) calculation formula is as follows:

[0109]

[0110] Function f(·,·) is the normalized pointwise mutual information, and its calculation formula is:

[0111]

[0112] Among them, P(w i ,w j ) is the term w i With the term w j The probability of co-occurrence in the same document, P(w i) is the term w i The marginal probability of occurrence, P(w j ) is the term w j The marginal probability of occurrence is selected. The number of topics K corresponding to the maximum value of the topic consistency index is determined as the parameter value for the final ETM model training.

[0113] For example, the training parameter settings are: learning rate is 0.001, training cycle is 150, batch size is 1000, and other parameters except the number of topics K use the default values ​​of the ETM model. The number of topics K is repeated in the interval [5,100] with a step size of 5 to obtain the topic consistency results corresponding to each experimental model, such as Figure 1 As shown:

[0114] When the number of topics is 50, the topic consistency reaches the maximum value and has better interpretability, so the number of topics K in the ETM training parameters is determined to be 50.

[0115] S62: The word embedding obtained in S4 and the technical documents recorded in the form of bag of words obtained in S5 are used as model inputs, and topic modeling is performed based on the ETM model parameter values ​​determined in S61 to obtain three output results in the form of matrices:

[0116] Topic-term probability distribution matrix β:

[0117]

[0118] Row vector β k· is the probability distribution of topic k (i.e., the kth topic among K topics) over all terms;

[0119] Document-topic probability distribution matrix θ:

[0120]

[0121] Row vector θ d· is the probability distribution of document d (i.e., the d-th document in D topics) on all topics;

[0122] Topic embedding matrix α:

[0123]

[0124] The column vector α·k is the topic embedding corresponding to topic k (i.e., the kth topic among K topics).

[0125] For example, the 300-dimensional word embeddings corresponding to the 19,709 terms obtained in Example S4 and the 67,135 technical documents recorded in the form of bag-of-words with a dictionary size of 9,042 obtained in Example S5 are used for topic modeling through the model parameter values ​​determined in Example S61, and finally the following is obtained:

[0126] Topic-term probability distribution matrix β:

[0127]

[0128] Document-topic probability distribution matrix θ:

[0129]

[0130] Topic embedding matrix α:

[0131]

[0132] The above results are presented with only three significant figures. The "topic" can be presented in two different forms through the obtained result matrix β and matrix α. When presented by matrix β, each topic can be represented by the term with the largest probability distribution value of the topic on the term, as shown in the following table:

[0133]

[0134]

[0135] The table shows some of the results, namely the 15 terms with the largest probability distribution values ​​for different topics on different terms. When displayed by α, the topic embedding corresponding to the topic in the embedding space can be visualized after dimensionality reduction, as well as the word embedding of the term closest to the cosine space distance of the topic embedding, such as Figure 2 shown.

[0136] It should be noted that the “topics” represent different technologies, and the technical contents represented by these technologies are represented by the terms with the largest probability distribution value on the topic, or by the terms corresponding to the word embeddings in the embedding space that are closest to the cosine space distance of the topic embedding.

[0137] Taking Topic 5 of the results in the figure as an example, the topic embedding corresponding to Topic 5 in the embedding space and the word embeddings corresponding to the 15 terms that are closest to the cosine space distance of the topic embedding are displayed. According to the meaning of the terms adjacent to Topic 5, it can be inferred that this topic represents the technology related to the temperature control and heating process in 3D printing technology.

[0138] S7: Use the cosine similarity calculation formula to measure the proximity between two topic embeddings, and obtain the proximity association matrix P that records the degree of proximity association between the topics;

[0139] The method of the present invention is based on the unique output of ETM - the topic embedding matrix α, which achieves a more accurate measurement of the degree of topic proximity. The existing methods for measuring the degree of topic proximity are all based on the topic-term probability distribution matrix β output by the traditional topic model. However, the traditional topic model only relies on the document set represented by the bag of words that can only capture the discrete relationship between terms during the generation process, which makes the topic-term probability distribution matrix β output by the model unable to capture the semantic content represented by the topic. Therefore, the degree of topic association calculated by this type of method is not accurate enough. The topic embedding matrix α output by the ETM model, based on the input word embedding data, realizes the expression of the semantic content contained in the topic in the form of a vector (i.e., embedding). The topic embedding can fully reflect the rich semantic information contained in the topic content, and further, a more accurate measurement of the degree of topic association at the semantic level can be obtained by a calculation method suitable for the degree of vector (i.e., embedding) relationship.

[0140] Measuring the proximity between two topic embeddings using the cosine similarity calculation formula includes the following steps:

[0141] S71: Calculate the element value in the proximity association matrix P based on the cosine similarity;

[0142]

[0143] P is a symmetric square matrix, and the calculation formula of element value is:

[0144]

[0145] Among them, α i is the topic embedding corresponding to topic i, α j is the topic embedding corresponding to topic j, ||α i || is the L-2 norm of the topic embedding corresponding to topic i, ||α j || is the L-2 norm of the topic embedding corresponding to topic j, p ij and p ji is the proximity correlation degree between topic i and topic j.

[0146] For example: Based on the results of the topic embedding matrix α obtained in Example S62, the proximity association matrix P is calculated using the cosine similarity calculation formula (proximity association is the first type of association in dual association);

[0147]

[0148] S72: scaling the value range of the element values ​​in the proximity association matrix P from [-1, 1] to [0, 1] by standardization processing, and the calculation formula for the standardization processing is:

[0149]

[0150] p′ ij is the element value of the standardized proximity correlation matrix P;

[0151] In the present invention, the standardization process is to ensure that the weights of the network edges subsequently constructed based on the proximity association matrix P are non-negative. Then, in order to ensure that there is no node self-connection in the network subsequently constructed based on the proximity association matrix P, the diagonal element values ​​of the proximity association matrix P are returned to 0, that is:

[0152] p′ ii =0

[0153] Finally, the processed proximity correlation matrix P is obtained.

[0154] For example, after normalizing the proximity correlation matrix P obtained in Example S71 and returning the diagonal elements to zero, the final proximity correlation matrix P is obtained:

[0155]

[0156] S8: Based on the document-topic probability distribution matrix θ, a combined association matrix C recording the combined association degree between topics is calculated by an association strength standardization method (combined association is the second type of association in dual association).

[0157] In the method of the present invention, a more accurate measurement of the degree of topic combination is achieved based on the association strength standardization method. The combination refers to the relationship between two different topics being used together, that is, the relationship of co-occurrence. Therefore, when measuring the degree of topic combination, it is necessary to use the measurement method of co-occurrence. The existing measurement methods of topic co-occurrence are all based on the document-topic probability distribution matrix θ. By setting a threshold value of the probability of topic occurrence or setting a threshold value of the number of topics contained in each document, θ is converted into a binary matrix and then a topic co-occurrence matrix is ​​constructed, thereby achieving the measurement of topic co-occurrence. This type of method will cause a large amount of topic probability information to be lost in the process of setting a threshold value for binary matrix conversion. In addition, the direct calculation of the topic co-occurrence degree based on the binary matrix will be affected by the size effect brought by the topic probability itself. The size effect will cause the topics that often appear in the document set to counterfactually have a high frequency of co-occurrence with other topics. If the size effect is not eliminated, the final topic co-occurrence result will not be able to reflect the true combination (i.e. co-occurrence) correlation degree between the topics. The association strength standardization method used in the steps of this method is a measurement method for measuring the deviation between the observed co-occurrence frequency and the expected co-occurrence frequency under the assumption of independence. Therefore, the associationstrength normalization method can eliminate the scale effect existing in co-occurrence data.

[0158] The combined association matrix C of the combined association degree between the record topics calculated by the association strength standardization method includes the following steps:

[0159] S81: Set a threshold η and set the element values ​​less than η in the document-topic probability distribution matrix θ to 0, thereby filtering rare events. The formula is expressed as:

[0160]

[0161] Explanation of reasons and necessity for filtering rare events:

[0162] Since the association strength normalization method may have calculation bias when facing rare events, in order to avoid the influence of rare events and obtain more accurate results, it is necessary to set a threshold for filtering rare events. The rare events are events with a very small probability of occurrence.

[0163] An example of calculation bias when the Association strength standardization method is used in the face of rare events: event A occurs 1 times, event B occurs 1 times, and event A and event B occur only once each, which is the case when event A and event B occur at the same time. The formula is: A = B = AB = 1. Then the probability of event A and event B occurring together calculated using the association strength standardization method is:

[0164]

[0165] Obviously, the probability value of event A and event B occurring together is calculated according to the association strength standardization method is a counterfactual, larger value.

[0166] For example, we set the threshold η to 0.01 and filter the rare events in the document-topic probability distribution matrix θ obtained in Example S62 based on η. The filtered matrix θ is:

[0167]

[0168] S82: The combined association matrix C of the combined association degree between the record topics is calculated using the association strength standardization method:

[0169]

[0170] C is a symmetric square matrix, and the calculation formula of element value is:

[0171]

[0172] θ id is the element value in the document-topic probability distribution matrix θ, that is, the probability that topic i appears in a document d; θ jd is the element value in the document-topic probability distribution matrix θ, that is, the probability that topic j appears in a document d; ∑ d θ id θ jd is the probability that topic i and topic j appear together in the document collection, ∑ d θ id is the total probability of topic i appearing in the document collection; ∑ d θ jd is the total probability of topic j appearing in the document collection; then, in order to ensure that there is no node self-connection in the network constructed based on the combined association matrix C, the diagonal element values ​​of the combined association matrix C are returned to 0, that is:

[0173] cii =0

[0174] Finally, the processed proximity correlation matrix C is obtained.

[0175] For example, based on the document-topic probability distribution matrix θ obtained in Example S81, the association strength normalization method is used to calculate and the diagonal element values ​​are normalized to 0 to obtain the combined association matrix C:

[0176]

[0177] S9: Based on the proximity correlation matrix P and the combined correlation matrix C, a dual network with two network levels is constructed And for dual network Visualize; identify dual networks using multiple centrality analysis methods based on multiple centrality indicators and factor analysis Dual core technology, core combination technology, core adjacent technology and edge technology in dual network In , each topic represents a technical subfield);

[0178] Dual Network From the hyperadjacency matrix express:

[0179]

[0180] Among them, I is the unit matrix, that is:

[0181]

[0182] In the super adjacency matrix The unit matrix I in the non-diagonal position has no practical meaning and is only used to record and identify the one-to-one correspondence between the same nodes in two network layers.

[0183] The present invention constructs a dual network The method steps can simultaneously embed multiple technology associations into a network-based model that includes different technologies and the associations between technologies, so as to study the complex associations between technologies. The existing network-based technology association analysis methods all use single-layer network models, so they are only suitable for exploring technology association analysis with only one type of association. However, the associations between different technologies are complex and include multiple types of associations. When faced with multiple associations, the single-layer network model can only aggregate different types of associations into one association for research, thereby losing the association information contained in different types of associations. The dual network can overcome this problem. The dual network contains different network levels, and the nodes in different network levels are the same, but the connection relationships between the nodes are different. Among them, the nodes represent different topics, and the connection relationships between the nodes represent the association relationships between the topics. Therefore, the use of the dual network can be used to represent two association relationships between the same group of technologies, thereby providing a basis for analyzing complex association patterns between technologies and identifying core technologies.

[0184] The multiple centrality analysis method steps based on multiple centrality indicators and factor analysis adopted by the present invention can help the innovation subject measure the relative importance of different technical sub-fields in the target technical field and identify core technologies. In the process of technological innovation, the recombination of knowledge elements in existing technical fields, or the reconfiguration of the combination of knowledge elements in technical fields that have been combined together, is an important way to give birth to new technical fields and new technical solutions. Therefore, exploring the correlation and correlation patterns between technologies, and judging the relative importance of different technologies play an important role in the discovery of technological opportunities. This method step can judge the relative importance of technologies and core technologies in a group of technologies with dual correlation types through a multiple centrality analysis method based on multiple centrality indicators and factor analysis, thereby discovering potential important technological opportunities.

[0185] The multiple centrality analysis method based on multiple centrality indicators and factor analysis used in the present invention is a method that is applicable across heterogeneous levels and can simultaneously consider multiple centrality indicators. As a structural feature, centrality can reveal the importance and influence of individual nodes in multiple networks. Existing multiple centrality analysis methods such as multiplex rank and versatility measure the importance of nodes by measuring the centrality of nodes in multiple networks. However, most of these methods are based on the assumption that networks at all levels in multiple networks represent similar types of relationships and have similar structures, and are not suitable for multiple networks with large differences in the types of relationships between network levels. In addition, the existing multiple centrality analysis methods can only consider one centrality indicator in each level, and cannot consider the centrality degree information reflected by different centrality indicators from different aspects. The multiple centrality analysis method based on multiple centrality indicators and factor analysis used in the present invention overcomes these two types of problems in the existing methods, is not only applicable to multiple networks containing heterogeneous levels, but also can simultaneously consider the position centrality degree represented by multiple centralities in each layer of the network, and realizes a more general and more accurate multiple centrality analysis method.

[0186] Specifically, the dual network It consists of two network levels. The nodes in the two network levels are the same, and the nodes represent the K topics obtained by S6. The links between the nodes within the two network levels respectively record the degree of association corresponding to the different association relationships between the same group of nodes, that is, the proximity association degree and the combined association degree between the K topics measured by S7 and S8 respectively. In addition to the links within the two network levels, there are also links between the two network levels. The link relationship between the two network levels records the one-to-one correspondence between the nodes representing the same node (i.e., the topic representing the technology) in the two network levels.

[0187] For example: Based on the proximity association matrix P obtained in example S72 and the combined association matrix C obtained in example S82, a dual network is constructed. From the hyperadjacency matrix It is expressed as:

[0188]

[0189] Then, for the dual network Specifically, the dual network constructed based on the Python toolkit pymnet Visualize the constructed dual network like Figure 3 shown.

[0190] Identifying dual networks using multiple centrality analysis methods based on multiple centrality indices and factor analysis The core technology steps are:

[0191] S91: Calculate the strength of each topic i in the proximity network layer among the K topics The calculation formula is: Calculate the strength of each topic i in the combined network layer among the K topics The calculation formula is: Using the Python toolkit NetworkX to calculate the dual network built by S9 The feature vector centrality of each topic i in the proximity network layer and the combination network layer in the K topics in and Current Betweenness Centrality Hehe Current proximity to centrality and

[0192] The current betweenness centrality and current closeness centrality used in this step can achieve more accurate betweenness centrality and closeness centrality in networks with technical associations. In networks with technical associations, the edges between each network node represent the propagation path of technical information between nodes. The simplest and most widely used closeness centrality and betweenness centrality calculations are based on the shortest path, but for networks with technical associations, technical information does not only flow and propagate along the shortest path. For example, in the dual network constructed by this method, In the network, the topics represented by the nodes consist of technical information. This technical information does not have a known ideal propagation path, so it will propagate in a random walk. In this type of network, the proximity and betweenness centrality measures based on the shortest path will produce incorrect results. To solve this problem, the method steps choose to obtain more accurate proximity and betweenness centrality measurements through current proximity centrality and current betweenness centrality applicable to information propagation process networks. These two current centrality measurement methods are still based on the same intuition as traditional proximity and betweenness centrality measurement methods, but more fully consider the process of information flowing and propagating in the network like current in the power grid.

[0193] Among them, the intensity Represents the sum of the associations between any topic i and other topics j in the K topics in the p network layer (i.e., the proximity network layer); pij is the association between topic i and topic j in the p network layer (i.e., the proximity network layer); higher strength This means that topic i is more closely related to other topics in the p network layer (i.e., the proximity network layer);

[0194] Among them, the intensity Represents the sum of the link weights between any topic i and other topics j in the K topics in the c network layer (i.e., the combined network layer); cij is the degree of association between topic i and topic j in the c network layer (i.e., the combined network layer); higher strength This means that topic i is more closely related to other topics in the c network layer (i.e., the combined network layer);

[0195] Among them, the eigenvector centrality function in the python toolkit NetworkX is used to obtain the eigenvector centrality of each topic i in the p network layer (i.e., the proximity network layer) corresponding to the proximity association matrix P. Eigenvector centrality The larger the value, the greater the influence of topic i in the p network layer (i.e., the proximity network layer);

[0196] Among them, the eigenvector centrality function in the Python toolkit NetworkX is used to obtain the eigenvector centrality of each topic i in the c network layer (i.e., the combined network layer) corresponding to the combined correlation matrix C Eigenvector centrality The larger the value, the greater the influence of topic i in the c network layer (i.e., the combined network layer);

[0197] Among them, the current betweenness centrality of each topic i in the p network layer (i.e., the proximity network layer) corresponding to the proximity association matrix P is obtained by the current_flow_betweenness_centrality function in the python toolkit NetworkX Current Betweenness Centrality The larger it is, the more important the topic i is for the p network layer (i.e., the proximity network layer).

[0198] Among them, the current betweenness centrality of each topic i in the c network layer (i.e., the combined network layer) corresponding to the combined correlation matrix C is obtained by the current_flow_betweenness_centrality function in the python toolkit NetworkX Current Betweenness Centrality (i) The larger the value, the more important the mediating role of topic i is for the c network layer (i.e., the combined network layer);

[0199] Among them, the current closeness centrality of each topic i in the p network layer (i.e., the proximity network layer) corresponding to the proximity association matrix P is obtained by the current_flow_closeness_centrality function in the python toolkit NetworkX Current proximity to centrality The larger the value, the closer the relationship between topic i and other topics is at the p network layer (i.e., the proximity network layer);

[0200] Among them, the current closeness centrality of each topic i in the c network layer (i.e., the combined network layer) corresponding to the combined correlation matrix C is obtained by the current_flow_closeness_centrality function in the python toolkit NetworkX Current proximity to centrality The larger the value, the closer the association between topic i and other topics is at the c network layer (i.e., the combined network layer);

[0201] Thus, the 8 centralities of each topic i among the K topics are calculated:

[0202] For example: the dual network constructed in the S9 example Eight centrality variables were calculated for each of the 50 topics i included in the dataset, and some of the results are as follows:

[0203]

[0204] S92: For each topic i among the K topics obtained in S91 As input data, we use the factor function of stata software to perform factor analysis (factor analysis is a data dimensionality reduction method that retains important information in the data set. Through factor analysis, multiple variables can be simplified into several unrelated factors that can explain the original variables). Finally, we get the centrality variables that can explain the four p network layers (proximity network layers) Commonality factor f p and the corresponding factor score f of each topic i in the K topics on the factor fp p (i), factor score f p The value of (i) can comprehensively reflect the centrality of the position of topic i in the p network layer (proximity network layer) represented by the four centrality variables, and can explain the centrality of the four centrality variables in the c network layer (combination network layer). Commonality factor f cAnd each topic i among the K topics has a factor f c The corresponding factor score f c (i), factor score f c The value of (i) can comprehensively reflect the centrality of the topic i represented by the four centrality variables in the c network layer (combined network layer); each topic i in the dual network The centrality in can be expressed as a vector Indicates that subject i is in the dual network The overall centrality of the network position.

[0205] For example, using the factor function of stata software, setting the parameters of the control factor analysis method to the principal component (pf), we perform factor analysis on the eight centrality variables calculated in the S91 example, and finally obtain the factor score f for each of the 50 topics i. p (i) and f c (i), the results are as follows:

[0206]

[0207]

[0208] By vector It is expressed as:

[0209]

[0210] S93: Setting factor score f for dividing topic i p (i) High and low thresholds γ p , set the factor score f of the partition topic i c (i) High and low thresholds γ c , which is used to judge the centrality of the network position of topic i in the proximity network layer and the combination network layer; specifically, the threshold γ in the proximity network layer p It can be set as the factor score f of K topics p (i) The median, third quantile, etc. of the combined network layer; the threshold γ c It can be set as the factor score f of K topics c (i) The median, third quantile, etc.

[0211] For example: set the threshold γ p Set to 50 topic factor scores f p (i), setting the threshold γc to the median of 50 topic factor scores f c The median of (i). Right now:

[0212] γ p=0.045,γ c =0.168

[0213] S94: factor score f p (i) As the x-axis, the factor score f c (i) As the y-axis, construct a two-dimensional coordinate system and use the vector obtained by S92 Construct a two-dimensional space based on the two thresholds γ set by S93 p and γ c Divide the two-dimensional space into four spatial regions to identify the dual network Dual core technology, core combination technology, core adjacent technology and edge technology in dual network In , each topic represents a technical subfield);

[0214] The dual core technologies refer to themes that show high centrality in both the proximity network layer and the combination network layer. The technical knowledge and foundations represented by these themes are generally related to the technologies represented by other themes, which can reduce the cost of technology combination and reflect higher combination potential.

[0215] The core combination technology refers to a topic with low centrality in the proximity network layer but high centrality in the combination network layer. This type of technology may be a technology with high knowledge barriers. Trying to learn such technology will increase learning and cognitive costs, so the technologies represented by other topics tend to directly combine such technologies with themselves to generate new solutions to technical problems. In addition, this type of technology may also be a new technology formed by widely combining other different technical elements.

[0216] The marginal technologies are topics that show low centrality in both the proximity network layer and the combination network layer. Such topics may be emerging technologies that are still in the early stages of development, focusing on more innovative and new theories and technologies that are not explored by other topics. In addition, such topics may also represent outdated technologies, so further in-depth evaluation may be required to determine whether they are still worth developing.

[0217] The core proximity technology refers to a topic with high centrality in the proximity network layer and low centrality in the combination network layer. Not all technical knowledge can be combined to solve new technical problems. For example, in biotechnology, not all protein pairs can be combined into compounds that effectively attack diseases. Therefore, in this case, it is necessary to integrate the technical knowledge of other topics and make certain improvements before it can be used to solve specific technical problems. Since the knowledge structure of the new technology formed after integration and improvement comes from other technologies to a certain extent, it will have a certain proximity to other related technologies.

[0218] For example: Based on the S92 example, we get a vector of 50 topic factor scores and the threshold γ set in the S93 example p and γ c , a two-dimensional space with four spatial regions is constructed, thereby distinguishing four types of technologies. The results are shown in Figure 4 . The first quadrant (factor score f p (i) High and factor score f c (i) high) is dual core technology; the second quadrant (factor score f p (i) Low and factor score f c (i) high) is a core combination technology; the third quadrant (factor score f p (i) Low and factor score f c (i) low) is marginal technology; the fourth quadrant (factor score f p (i) High and factor score f c (i) Low) is the core proximity technology. Each sample point in the two-dimensional space is the vector corresponding to a topic. Different topics are identified in the figure using the label style i.

[0219] Regarding the embodiments of the present invention, it should also be noted that, in the absence of conflict, the embodiments of the present invention and the features therein may be combined with each other to obtain new embodiments.

[0220] The above is only a preferred embodiment of the present invention, and does not limit the present invention in any form. The protection scope of the present invention shall be subject to the protection scope of the claims. Although the present invention has been disclosed as a preferred embodiment as above, it is not used to limit the present invention. Any technician familiar with this profession can make some changes or modifications to equivalent embodiments of equivalent changes using the technical content disclosed above without departing from the scope of the technical solution of the present invention. However, any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention still falls within the scope of the technical solution of the present invention.

Claims

1. A core technology identification method based on dual technology relevance, characterized by: The following steps are involved: Extract technical keywords based on technical documents, construct a technical search formula, and obtain D original technical documents from the technical database based on the technical search formula; Performing sentence segmentation and a first text preprocessing on the text content in each of the D original technical documents to obtain preprocessed technical document sentences, and training a word embedding model based on the preprocessed technical document sentences to obtain word embeddings corresponding to the terms contained in the D original technical documents; Performing a second text preprocessing based on the D original technical documents to obtain technical documents recorded in the form of word bags; Based on the word embedding and the technical documents in the bag-of-words form, topic modeling is performed through the ETM model to obtain three results, including a topic-term probability distribution matrix β, a document-topic probability distribution matrix θ, and a topic embedding matrix α; Based on the topic embedding matrix α, the cosine similarity calculation formula is used to measure the proximity between two topic embeddings, and the proximity association matrix P that records the proximity association strength between topics is obtained. Based on the document-topic probability distribution matrix θ, the combined association matrix C that records the combined association strength between topics is calculated by the association strength standardization method. Based on the proximity association matrix P and the combined association matrix C, a dual network with two network levels is constructed. And for dual network Visualize; identify dual networks using multiple centrality analysis methods based on multiple centrality indicators and factor analysis Dual-core technology, core combination technology, core adjacent technology and edge technology in the industry; Dual Network From the hyperadjacency matrix express: in, The dual core technologies refer to topics that show high centrality in network positions both in the proximity network layer and in the combined network layer; The core combination technology refers to a topic with low centrality in the proximity network layer and high centrality in the combination network layer; The edge technologies refer to topics that show low centrality in both the proximity network layer and the combination network layer; The core proximity technique refers to topics that have high centrality in the proximity network layer and low centrality in the combination network layer.

2. The core technology identification method based on dual technology relevance according to claim 1 is characterized by: The word embedding model is a Word2Vec model or a glove model.

3. The core technology identification method based on dual technology relevance according to claim 1 is characterized in that: The specific steps of sentence segmentation and the first text preprocessing include: The text content of each original technical document is segmented into technical document sentences with sentences as units through the Python toolkit NLTK. The python toolkit spaCy is used to perform the first text preprocessing on the technical document after sentence segmentation to obtain the preprocessed technical document sentences.

4. The core technology identification method based on dual technology relevance according to claim 3 is characterized by: The first text preprocessing process is as follows: first, numbers, punctuation marks and words with a character length less than 3 in the text are deleted; then, the remaining words are restored to their parts of speech and converted to lowercase; finally, sentences whose sentence length is less than the "window size" parameter value set by the subsequent Word2Vec model are deleted to obtain the preprocessed technical document sentences.

5. The core technology identification method based on dual technology relevance according to claim 1 is characterized in that: The second text preprocessing uses the python toolkit spaCy. The processing process includes deleting numbers, punctuation marks, stop words, terms with a character length of less than 3, and non-noun terms in the original technical documents, restoring the remaining terms and converting them to lowercase, deleting terms with a document frequency of less than 3 in the technical document collection, and finally obtaining the preprocessed technical documents recorded in the form of a bag of words with a dictionary size of V.

6. The core technology identification method based on dual technology relevance according to claim 1 is characterized by: Based on the word embedding and the technical documents in the form of bag of words, the topic modeling process is performed through the ETM model to obtain three results, including the topic-term probability distribution matrix β, the document-topic probability distribution matrix θ and the topic embedding matrix α. The specific steps are as follows: Based on the word embedding and the technical documents recorded in the bag-of-words form as input data for training the ETM model, the optimal setting value of the parameter of the number of topics K in the ETM model is determined through repeated experiments; The technical documents recorded in the word embedding and bag-of-words form are used as model inputs, and topic modeling is performed based on the ETM model parameter value topic number K to obtain three matrix output results: Topic-term probability distribution matrix β: Row vector β k· is the probability distribution of topic k on all terms, where k is the kth topic among K topics; Document-topic probability distribution matrix θ: Row vector θ d· is the probability distribution of document d on all topics, where d is the d-th document in D topics; Topic embedding matrix α: The column vector α·k is the topic embedding corresponding to topic k, where k is the kth topic among K topics.

7. The core technology identification method based on dual technology relevance according to claim 1 is characterized by: The specific steps of using the cosine similarity calculation formula to measure the proximity between two topic embeddings and obtain the proximity association matrix P that records the proximity association strength between topics are as follows: Use the cosine similarity calculation formula to obtain the element values ​​in the proximity association matrix P; P is a symmetric square matrix, and the calculation formula of element value is: Among them, α i is the topic embedding corresponding to topic i, α j is the topic embedding corresponding to topic j, ||α i || is the L-2 norm of the topic embedding corresponding to topic i, ||α j || is the L-2 norm of the topic embedding corresponding to topic j, p ij and p ji is the proximity correlation degree between topic i and topic j; The range of element values ​​in the proximity association matrix P is scaled from [-1, 1] to [0, 1] by standardization. The standardization calculation formula is: p′ ij is the normalized matrix element value; The diagonal element values ​​of the proximity association matrix P are returned to 0, that is: p′ ii ＝0 Finally, the processed proximity correlation matrix P is obtained.

8. The core technology identification method based on dual technology relevance according to claim 1 is characterized in that The specific steps of calculating the combined association matrix C that records the combined association strength between topics based on the document-topic probability distribution matrix θ by the association strength standardization method are as follows: Set a threshold η and set the element values ​​less than η in the document-topic probability distribution matrix θ to 0, thereby filtering rare events. The formula is expressed as: The combined association matrix C of the combined association strength between the recorded topics is calculated using the association strength standardization method: C is a symmetric square matrix, and the calculation formula of element value is: θ id is the element value in the document-topic probability distribution matrix θ, that is, the probability that topic i appears in document d; θ jd is the element value in the document-topic probability distribution matrix θ, that is, the probability that topic j appears in a document d; ∑ d θ id θ jd is the probability that topic i and topic j appear together in the document collection, ∑ d θ id is the total probability of topic i appearing in the document collection; ∑ d θ jd is the total probability of topic j appearing in the document collection; The diagonal element values ​​of the combined correlation matrix C are returned to zero, and finally the processed combined correlation matrix C is obtained. The return-to-zero processing formula is: c ii ＝0。 9. The core technology identification method based on dual technology relevance according to claim 1 is characterized in that: Dual networks were identified using a multi-centrality analysis method based on multiple centrality indices and factor analysis. The steps of dual core technology, core combination technology, core adjacent technology and edge technology in the process include: Calculate the strength of each topic i in the proximity network layer among the K topics The calculation formula is: Calculate the strength of each topic i in the combined network layer among the K topics The calculation formula is: The dual network is calculated using the python toolkit NetworkX The feature vector centrality of each topic i in the proximity network layer and the combination network layer in the K topics in and Current Betweenness Centrality and Current proximity to centrality and Among them, the eigenvector centrality of each topic i in the p network layer corresponding to the proximity association matrix P is obtained through the eigenvector_centrality function in the python toolkit NetworkX The eigenvector_centrality function in the Python toolkit NetworkX is used to obtain the eigenvector centrality of each topic i in the c network layer corresponding to the combined correlation matrix C. The current betweenness centrality of each topic i in the p network layer corresponding to the proximity association matrix P is obtained through the current_flow_betweenness_centrality function in the python toolkit NetworkX The current betweenness centrality of each topic i in the c network layer corresponding to the combined correlation matrix C is obtained through the current_flow_betweenness_centrality function in the python toolkit NetworkX The current closeness centrality of each topic i in the p network layer corresponding to the proximity association matrix P is obtained through the current_flow_closeness_centrality function in the Python toolkit NetworkX The current closeness centrality of each topic i in the c network layer corresponding to the combined association matrix c is obtained through the current_flow_closeness_centrality function in the Python toolkit NetworkX For each topic i among the K topics As input data, factor analysis was performed using the factor function of stata software to obtain the centrality variables that can explain the four layers of the proximity network. Commonality factor f p And each topic i among the K topics has a factor f p The corresponding score f p (i) and the centrality variables that can explain the four combinations of network layers Commonality factor f c and the corresponding factor score f of each topic i in the K topics on the factor fc c (i), each subject i in the dual network The centrality in is represented as a vector Set the factor score f for the partition topic i p (i) High and low thresholds γ p , set the factor score f of the partition topic i c (i) High and low thresholds γ c , which is used to judge the centrality of the network position of subject i in the proximity network layer and the combination network layer respectively; The factor score f p (i) As the x-axis, the factor score f c (i) As the y-axis, a two-dimensional coordinate system is constructed, and based on the vector Construct a two-dimensional space based on two thresholds γ p and γ c Divide the two-dimensional space into four spatial regions to identify the dual network Dual-core technology, core combination technology, core adjacent technology and edge technology.

10. The core technology identification method based on dual technology relevance according to claim 1 is characterized in that: The threshold γ p Set to the median or third quantile of the factor scores of K topics in the proximity network layer; the threshold γ c Set to the median or third quantile of the factor scores of the K topics in the combined network layer.

Images