A dual-network-based dual-technology correlation analysis method

Abstract

The application discloses a dual-network-based dual-technology correlation analysis method, which can embed two types of technology correlations of proximity and combination in a network system, visualize the correlation structure between technologies, and analyze various correlation modes between technologies. First, relevant technical documents are collected, and the text content in the documents is preprocessed. Second, the word embedding corresponding to the word items contained in the technical documents is obtained through Word2Vec model training. Third, the obtained word embedding and the preprocessed technical documents are taken as inputs, and the topics representing different technologies are extracted through an ETM-based topic modeling process. Fourth, the proximity and combination correlation degrees between the topics are measured by using the cosine similarity and association strength standardization method. Fifth, the topics and the proximity and combination correlation between the topics are taken as nodes and node links in the network respectively to construct a dual network. Finally, a two-dimensional space is constructed based on multiple links to distinguish four correlation modes between technologies.

Description

TECHNICAL FIELD

[0001] The present application relates to a dual technology correlation analysis method based on dual networks. BACKGROUND

[0002] Technology correlation shapes the technology diversification pattern of innovation subjects and influences the development direction of technological innovation. Clarifying the correlation between technologies helps to clarify the relative importance of different technologies and the correlation structure in the technology field. Exploring the correlation between different technologies can help innovation subjects to comprehensively grasp the technology dynamics and assist in making implementation decisions for technological innovation, thereby promoting the development of new productivity.

[0003] Existing research ignores the multiple correlation types of technology correlation and only considers a single relationship type between technologies. However, the concept of technology correlation is not single-dimensional and contains multiple correlation types. Different types of correlation have different degrees of correlation between the same group of technologies and exhibit complex correlation patterns, which contain more rich technology information compared to only considering a single correlation type. Considering the multiple types of technology correlation can help reveal the real relationship structure between multiple technology fields in the technology system, enabling innovation subjects to make correct technology innovation decisions. Therefore, a method is needed that can simultaneously embed multiple correlation types of technology correlation to conduct more comprehensive and accurate correlation pattern and correlation structure research.

[0004] Current research methods rely on single-layer networks to study the 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 correlation types. When using such methods to study multiple correlations, only one type of correlation can be aggregated, but this will lose the correlation information contained between different correlation types. Therefore, a method is needed that can simultaneously study and reveal the multiple correlation relationships between technologies.

[0005] The measurement of the degree of correlation between technologies is the basis for studying technology correlation, which determines the correlation structure between technologies. Although some existing methods use keyword-based text mining techniques to attempt to use text data to achieve more accurate correlation degree measurement, keyword-based methods can only capture the discrete relationships between keywords and ignore semantic connections. In addition, such methods cannot handle the frequent occurrence of technical terms, synonyms, and polysemous words in patent text. Therefore, a method is needed that is based on text and takes into account the semantics of the text to achieve more accurate correlation degree measurement. SUMMARY

[0006] The purpose of the present application is to provide a dual technology correlation analysis method based on dual networks to solve the technical problems raised in the background art, comprising the following steps:

[0007] extracting technology keywords based on the technical document, constructing a technology search formula, and obtaining D original technical documents from a technical database based on the technology search formula;

[0008] performing sentence segmentation and first text preprocessing on the text content in the D original technical documents to obtain preprocessed technical document sentences, and obtaining word embeddings corresponding to the word items contained in the D original technical documents based on the preprocessed technical document sentences through a training process of a word embedding model;

[0009] performing second text preprocessing based on the D original technical documents to obtain technical documents recorded in the form of a bag of words;

[0010] performing topic modeling through an ETM model based on the word embeddings and the technical documents in the form of a bag of words to obtain three results, including a topic-word probability distribution matrix β, a document-topic probability distribution matrix θ, and a topic embedding matrix α;

[0011] measuring the proximity between two topic embeddings based on the topic embedding matrix α using a cosine similarity calculation formula to obtain a proximity association matrix P recording the proximity association degree between topics, calculating a combination association matrix C recording the combination association degree between topics based on the document-topic probability distribution matrix θ through an association strength standardization method, and constructing a dual network with two network levels based on the proximity association matrix P and the combination association matrix C performing visualization on the dual network constructing a two-dimensional space based on the multiple link vectors recording the association degree between topic pairs in the dual network p and λ c ; dividing the two-dimensional space into four different space regions based on the threshold values λ p and λ c , determining four association modes of topic pairs, including dual association, proximity dominant association, combination dominant association, and weak association;

[0012] The dual network can be represented by a super adjacency matrix

[0013]

[0014] wherein

[0015] ​​​

[0016] Further, the word embedding model is a Word2Vec model or a Glove model.

[0017] Further, the specific steps of the sentence segmentation and the first text preprocessing include:

[0018] The text content in each original technical document is segmented into technical document sentences by using the python toolkit NLTK.

[0019] The technical document sentences are subjected to the first text preprocessing by using the python toolkit spaCy to obtain preprocessed technical document sentences.

[0020] Further, the first text preprocessing process: first, delete the numbers, punctuation marks and word items with a character length less than 3 in the text; then, restore the part of speech of the remaining word items and convert them to lowercase; finally, delete the sentences with a length less than the "window size" parameter value set by the subsequent Word2Vec model, to obtain the preprocessed technical document sentences.

[0021] Further, the second text preprocessing uses the python toolkit spaCy, and the processing process includes deleting the numbers, punctuation marks, stop words, word items with a character length less than 3 and non-nominal word items in the original technical document, restoring the part of speech of the remaining word items and converting them to lowercase, deleting the word items with a document frequency less than 3 in the technical document set, and finally obtaining the preprocessed technical document recorded in the form of a bag of words with a dictionary size of V.

[0022] Further, based on the word embedding and the technical document recorded in the form of a bag of words, theme modeling is performed by using the ETM model to obtain three results, including a theme-word item probability distribution matrix β, a document-theme probability distribution matrix θ and a theme embedding matrix α.

[0023] The specific steps are as follows:

[0024] Based on the word embedding and the technical document recorded in the form of a bag of words as input data for training the ETM model, the optimal setting value of the theme number K parameter in the ETM model is determined by repeated experiments;

[0025] Based on the word embedding and the technical document recorded in the form of a bag of words as model input, theme modeling is performed based on the theme number K parameter value of the ETM model to obtain three matrix-form output results:

[0026] Theme-word item probability distribution matrix β:

[0027]

[0028] Row vector β k· is the probability distribution of topic k over all terms, where k is the kth topic among K topics;

[0029] Document-topic probability distribution matrix θ:

[0030]

[0031] Row vector θ d· is the probability distribution of document d over all topics, where d is the dth document among D documents;

[0032] Topic embedding matrix α:

[0033]

[0034] Column vector α·k is the topic embedding corresponding to topic k, where k is the kth topic among K topics.

[0035] Further, the proximity correlation matrix P recording the proximity correlation degree between topics is obtained by measuring the proximity between two topic embeddings using the cosine similarity calculation formula, and the specific steps are as follows:

[0036] The element value in the proximity correlation matrix P is obtained by using the cosine similarity calculation formula;

[0037]

[0038] P is a symmetric matrix, and the calculation formula of the element value is:

[0039]

[0040] where 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 are the proximity correlation degree between topic i and topic j;

[0041] The value range of the element value in the proximity correlation matrix P is scaled from [-1, 1] to [0, 1] by standardization processing, and the standardization processing calculation formula is:

[0042]

[0043] p′ ij is the matrix element value after standardization;

[0044] The diagonal elements of the proximity association matrix P are processed by zeroing, that is:

[0045] p′ ii = 0

[0046] The processed proximity association matrix is finally obtained.

[0047] Further, the specific steps for calculating the combination association matrix C representing the combination association degree between record topics based on the document-topic probability distribution matrix θ by the association strength standardization method are as follows:

[0048] A threshold η is set, and the element values in the document-topic probability distribution matrix θ smaller than η are set to 0, thereby filtering rare events, and the formula is expressed as:

[0049]

[0050] The combination association matrix C representing the combination association degree between record topics is calculated by the association strength standardization method:

[0051]

[0052] C is a symmetric matrix, and the calculation formula of the element value is:

[0053]

[0054] θ id is the element value in the document-topic probability distribution matrix θ, that is, the probability of the occurrence of topic i in document d; θ jd is the element value in the document-topic probability distribution matrix θ, that is, the probability of the occurrence of topic j in document d; ∑ d θ id θ jd is the probability of the co-occurrence of topic i and topic j in the document set; ∑ d θ id is the total probability of the occurrence of topic i in the document set; ∑ d θ jd is the total probability of the occurrence of topic j in the document set;

[0055] The diagonal elements of the combination association matrix C are processed by zeroing, and the processed combination association matrix C is finally obtained, and the zeroing processing formula is:

[0056] c ii = 0.

[0057] Further, based on the double network The multi-link vector that records the degree of association between topic i and topic j Steps to construct a two-dimensional space:

[0058] In dual network Construct multiple link vectors between any topic i and topic j p ij with c ij are the association degree of topic i and topic j in the proximity network layer and the association degree in the combination network layer, respectively, where p ij =p ji ,c ij =c ji ;

[0059] Based on the multiple linkage vector The two dimensional features in construct a two-dimensional space, where The first element p in is used as the x-axis, and the second element c is used as the y-axis.

[0060] Furthermore, the threshold θ p Set to the median, third quantile, or 80th quantile of the association degree value in the proximity network layer; the threshold λ c Fixed to 1.

[0061] 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, while visualizing the association structure between technologies and analyzing various association patterns between technologies. First, relevant technical documents are collected and the text content in the documents is preprocessed; secondly, the word embeddings corresponding to the terms contained in the technical documents are obtained through Word2Vec model training; then, the obtained word embeddings and the preprocessed technical documents are used as input, and topics representing different technologies are extracted through ETM-based topic modeling; then, the cosine similarity and association strength standardization methods are used to measure the proximity and combination association degree between topics; next, the proximity and combination association between topics are used as nodes and node links in the network to construct a dual network. Finally, a two-dimensional space is constructed based on multiple links to distinguish the four association patterns between technologies. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. 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 any creative work.

[0063] Figure 1 It is a topic consistency display diagram of the ETM model corresponding to different number of topics K;

[0064] Figure 2 This is a graph showing the topic embedding of topic 5 and the word embedding of the 15 terms closest to its cosine space;

[0065] Figure 3 It is a dual network display diagram embedded with dual technology correlation;

[0066] Figure 4 It is a two-dimensional spatial display diagram based on multiple links;

[0067] Figure 5 It is a two-dimensional spatial display diagram that distinguishes four types of association patterns.

[0068] 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

[0069] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. 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 making creative efforts are within the scope of protection of the present invention.

[0070] The present invention provides a dual technology correlation analysis method based on a dual network, comprising the following steps:

[0071] S1: Extract technical keywords based on technical documents and construct technical search formulas;

[0072] For example, the technical documents are multiple articles related to the field of 3D printing technology. Keywords describing this technology are extracted from these documents, and a technical search formula suitable for the technology database is constructed. For example, a specific search formula suitable for the incoPat patent database is as follows:

[0073] “TIAB=(3D printing) OR TIAB=(three-dimensional printing) OR TIAB=(3DP) OR TIAB=(additive manufacturing) OR TIAB=(rapid prototyping) OR TIAB=(three-dimensional manufacturing)”

[0074] S2: Obtain D original technical documents from the technical database based on the technical search formula;

[0075] For example, the technical database is a patent database (incoPat patent database or Derwent Innovations Index patent database), and for example, based on the technical search formula for 3D printing technology constructed in S1, 67,231 original technical documents are searched from the incoPat patent database, and preferably, the content of the original technical documents is the patent specification title text and the abstract text based on which the subsequent method analysis is performed.

[0076] S3: performing sentence segmentation and first text preprocessing on the text content of the D original technical documents to obtain preprocessed technical document sentences;

[0077] Specifically, the specific steps of sentence segmentation and first text preprocessing include:

[0078] S31: using the python toolkit NLTK to perform segmentation on the text content of each original technical document to obtain technical document sentences in the form of sentences;

[0079] For example, the text content of the 67,231 original technical documents obtained in S2 is segmented to obtain 590,634 technical document sentences.

[0080] S32: using the python toolkit spaCy to perform first text preprocessing on the technical document sentences after sentence segmentation to obtain preprocessed technical document sentences.

[0081] Specifically, the first text preprocessing process includes the following steps: first, deleting numbers, punctuation marks, and word items with a character length less than 3 in the text; then, performing part-of-speech restoration on the remaining word items and converting them to lowercase; finally, deleting sentences with a length (the number of word items contained in the sentence) less than the “window size” parameter value set by the subsequent Word2Vec model to obtain preprocessed technical document sentences.

[0082] For example, the 590,634 technical document sentences obtained in S31 are preprocessed to obtain 561,733 preprocessed technical document sentences.

[0083] S4: based on the preprocessed technical document sentences obtained in S3, using the python toolkit Gensim to train a Word2Vec model to obtain word embeddings corresponding to the word items contained in the D original technical documents; the word embedding result will be used as one of the input data for the subsequent ETM model training.

[0084] For example, when training the Word2Vec model, the parameters are set as follows: the model architecture is Skip-Gram, the number of negative samples is 15, the dimension of word embeddings 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 are the default parameter values of the Word2Vec model API of Gensim. Based on the 561733 technical document sentences obtained in S32, the 300-dimensional word embeddings corresponding to the 19709 word items are obtained through the Word2Vec model training with the above parameters.

[0085] Preferably, in the present application, the word embedding model is a Word2Vec model, and of course, in addition to the above model, it can also be other word embedding models such as Glove.

[0086] Firstly, the Word2Vec model is simple and efficient, 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 the case of limited computing resources, which makes it unnecessary for researchers and practitioners to use a large amount of hardware support.

[0087] Secondly, the Word2Vec model has a fast training speed. The Word2vec model uses some acceleration methods such as Negative Sampling and Hierarchical Softmax, which makes it faster in training.

[0088] Finally, the Word2Vec model supports incremental training. This means that the Word2Vec model can be updated with new data without starting from scratch, which is a very useful feature for applications that need to constantly introduce new information.

[0089] S5: Perform second text preprocessing on the D original technical documents described in S2 to obtain technical documents recorded in the form of a bag of words; the technical documents recorded in the form of a bag of words will be one of the input data for subsequent ETM model training;

[0090] Specifically, the second text preprocessing uses the python toolkit spaCy, and the processing procedure includes deleting numbers, punctuation marks, stop words, word items with a character length less than 3 and non-noun word items in the original technical documents. The stop words refer to word items that frequently appear in a language and contribute less to the actual semantics of the text, such as articles, prepositions and the like. Finally, the remaining word items are subjected to part-of-speech reduction and converted into lowercase, and word items with a document frequency less than 3 in the technical document set are deleted. Finally, the technical documents in the form of a bag of words with a dictionary size of V are obtained after preprocessing. The dictionary refers to the set of all unique (i.e. unique, non-repeated) word items extracted from the technical documents.

[0091] For example, the 67,231 original technical documents obtained in the S2 example are subjected to the second text preprocessing, and 67,135 technical documents in the form of a bag of words with a dictionary size of 9,042 are obtained after preprocessing.

[0092] During the second text preprocessing, the selection of only retaining the word items of the noun part-of-speech can improve the training efficiency of the topic model and improve the semantic consistency and interpretability of the topic results. The components and terms related to methods, tools and the like are usually described in the form of nouns, so the word items of other parts-of-speech will become noise during the training of the topic model, thereby affecting the training efficiency of the model and the accuracy of the results, and cannot play a positive role in the meaning explanation of the topic results. Therefore, during the second text preprocessing, the word items other than the noun part-of-speech are deleted to solve these problems and play a positive role in promoting the training process of the topic model.

[0093] S6: Based on the word embedding obtained in S4 and the technical documents in the form of a bag of words obtained in S5, the topic modeling is performed through the ETM model, and three results are obtained, including a topic-word item probability distribution matrix β, a document-topic probability distribution matrix θ and a topic embedding matrix α;

[0094] Firstly, the ETM used in the present application has higher running efficiency and computing efficiency. Traditional topic models such as LSA (Latent Semantic Allocation) and LDA (Latent Dirichlet allocation) have lower computing efficiency and running efficiency. The traditional topic model has a parameter inference process specific to different models, and the complexity of the parameter inference process increases with the complexity of the model itself. In addition, the traditional topic model consumes a large amount of time when facing large document sets. The ETM model used in the method steps is a neural network-based topic model, so the model parameters can be inferred through the automatic gradient backpropagation process, and parallel computing is allowed, thereby greatly improving the computing efficiency of the topic modeling process.

[0095] 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 embeddings 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.

[0096] Furthermore, the ETM model used is robust to stop words and rare words. Traditional topic models reduce the interpretability of their results when encountering stop words and rare words. This is because stop words and rare words appear within the content of topic results, making them unhelpful for interpreting them. However, the ETM model used in this method generates separate topics for meaningless and frequently occurring stop words and rare words, distinguishing meaningful terms from these, thereby improving the interpretability of the model results.

[0097] Finally, thanks to the input word embedding data, the ETM model used has unique outputs that other topic models do not have.

[0098] Topic embedding. Traditional topic models only have two outputs: the topic-term probability distribution matrix β and the document-topic probability distribution matrix θ. ETM, on the other hand, has a unique output: the topic embedding matrix α. Topic embeddings express the semantic content and rich contextual semantic information of a topic through vectors (i.e., embeddings), providing a foundation for more accurate measurement of the degree of association between topics.

[0099] The process of obtaining three result matrices through topic modeling using the ETM model includes the following steps:

[0100] S61: The word embeddings obtained in S4 and the technical documents recorded in the bag-of-words format obtained in S5 are used as input data for training the ETM model. The optimal setting value of the parameter K for the number of topics in the ETM model is determined through repeated experiments.

[0101] For example, the other parameters in the ETM model remain unchanged, set different topic quantity K for repeated experiments, and calculate the topic coherence (TC) index of each experimental model result. The higher the topic coherence value, the stronger the interpretability of the topic model. The topic coherence (TC) calculation formula is as follows:

[0102]

[0103] The function f(·,·) is the standard point mutual information, and the calculation formula is as follows:

[0104]

[0105] Where, P(w i ,w j ) is the probability of the word w i and the word wjco-occur in the same document, P(w i ) is the marginal probability of the word w i , and P(w j ) is the marginal probability of the word w j . The topic quantity K corresponding to the maximum value of the topic coherence index is selected as the final parameter value of the ETM model training.

[0106] For example, the training parameters are set as follows: the learning rate is 0.001, the training period is 150, the batch size is 1000, and the other parameters not mentioned except the topic quantity K use the default values of the ETM model. The topic quantity K is repeatedly experimented in the interval [5, 100] with a step of 5, and the corresponding topic coherence results of each experimental model are obtained, as shown in Figure 1 .

[0107] When the topic quantity is 50, the topic coherence reaches the maximum value, and has better interpretability. Therefore, the topic quantity K in the ETM training parameters is determined as 50.

[0108] S62: Take the word embedding obtained in S4 and the technical document recorded in the form of bag of words obtained in S5 as model input, and perform topic modeling based on the ETM model parameter value determined in S61 to obtain three matrix form output results:

[0109] Topic-word probability distribution matrix β:

[0110]

[0111] The row vector β k· is the probability distribution of the topic k (i.e. the kth topic in the K topics) on all word items.

[0112] Document-topic probability distribution matrix Θ:

[0113]

[0114] Row vector θd d· is the probability distribution of document d (i.e., the d-th document among D documents) over all topics;

[0115] Topic embedding matrix α:

[0116]

[0117] Column vector α·k is the topic embedding of topic k (i.e., the k-th topic among K topics).

[0118] As an example, given the 300-dimensional word embeddings of 19,709 word items obtained in the S4 example and the 67,135 technical documents in bag-of-words form with a dictionary size of 9,042 obtained in the S5 example, the model parameters determined in the S61 example are used to perform topic modeling, and finally we obtain:

[0119] Topic-word probability distribution matrix β:

[0120]

[0121] Document-topic probability distribution matrix Θ:

[0122]

[0123] Topic embedding matrix α:

[0124]

[0125] The above results are shown with only three significant digits. The "topics" can be displayed in two different forms by the obtained result matrices β and α, respectively. When displayed by matrix β, each topic can be represented by the word item with the largest probability distribution value of the topic on the word item, as shown in the following table:

[0126]

[0127]

[0128] The figure shows part of the results, i.e., the 15 word items with the largest probability distribution values of different topics on different word items. When displayed by α, the topic embedding corresponding to the topic in the reduced dimension visualization embedding space and the word embedding of the word item closest to the cosine space distance of the topic embedding can be visualized, as shown in Figure 2 .

[0129] It should be noted that the "topic" represents different technologies, and the content of the technology represented by the technology is represented by the word item with the maximum probability value on the topic or the word embedding corresponding to the word embedding with the closest cosine space distance to the topic embedding in the embedding space.

[0130] Taking the topic 5 of the result in the figure as an example, the topic embedding corresponding to the topic 5 in the embedding space and the word embedding corresponding to the 15 nearest word items of the topic embedding are shown. According to the meaning of the word items adjacent to the topic 5, it can be inferred that the topic represents the technology related to temperature control and heating process in 3D printing technology.

[0131] S7: Measure the proximity between two topic embeddings using the cosine similarity calculation formula to obtain the proximity association matrix P recording the proximity association degree between topics;

[0132] The method of the present application realizes more accurate measurement of the proximity degree of the topic based on the unique output of ETM, the topic embedding matrix a. The existing measurement methods of the proximity degree of the topic are all based on the topic-word item probability distribution matrix b output by the traditional topic model. However, the traditional topic model only relies on the bag-of-words representation of the document set during the generation process, which can only capture the discrete relationship between word items, which makes the topic-word item probability distribution matrix b output by the model unable to capture the semantic content represented by the topic. Therefore, the association degree of the topic calculated by this kind of method is not accurate enough. The topic embedding matrix a output by the ETM model realizes the expression of the semantic content contained in the topic in the form of vector (i.e. embedding) based on the input word embedding data. The topic embedding can fully reflect the rich semantic information contained in the topic content, and further through the calculation method suitable for the relationship degree of vector (i.e. embedding), the more accurate measurement of the association degree of the topic at the semantic level can be obtained.

[0133] Measuring the proximity between two topic embeddings using the cosine similarity calculation formula includes the following steps:

[0134] S71: Calculate the element value in the proximity association matrix P based on the cosine similarity;

[0135]

[0136] P is a symmetric matrix, and the calculation formula of the element value is:

[0137]

[0138] Wherein, a i is the topic embedding corresponding to the topic i, a j is the topic embedding corresponding to the topic j, and ||a 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 association degree between topic i and topic j.

[0139] For example, based on the results of the topic embedding matrix α obtained in the S62 example, the proximity association matrix P (the proximity association is the first type of association in the double association) is calculated using the cosine similarity calculation formula:

[0140]

[0141] S72: The value range of the element values in the proximity association matrix P is scaled from [-1, 1] to [0, 1] through standardization processing, and the standardization processing calculation formula is:

[0142]

[0143] p′ ij is the element value of the standardized proximity association matrix P;

[0144] In the present application, the standardization processing is to ensure that the weight of the network edge constructed based on the proximity association matrix P in the subsequent is non-negative. Then, in order to ensure that there is no node self-connection in the network constructed based on the proximity association matrix P in the subsequent, the diagonal element value of the proximity association matrix P is processed to zero, that is:

[0145] p′ ii = 0

[0146] The processed proximity association matrix P is finally obtained.

[0147] For example, after standardization and diagonal element processing of the proximity association matrix P obtained in the S71 example, the final proximity association matrix P is obtained:

[0148]

[0149] S8: Based on the document-topic probability distribution matrix θ, the combination association matrix C (combination association is the second type of association in the double association) of the combination association degree between the recorded topics is calculated through the association strength standardization method.

[0150] In the method, the association strength standardization method is used to measure the combination degree of topics more accurately. The combination refers to the relationship that two different topics are combined for use, i.e., the co-occurrence relationship. Therefore, the co-occurrence degree measurement method is needed when measuring the combination degree of topics. The existing methods for measuring the co-occurrence degree of topics are based on the document-topic probability distribution matrix θ. By setting a threshold for the probability of topic occurrence or setting a threshold for the number of topics contained in each document, the θ is converted into a binary matrix to construct a topic co-occurrence matrix, thereby realizing the measurement of the co-occurrence degree of topics. In the process of setting the threshold for binary matrix conversion, a large amount of topic probability information is lost. In addition, the direct calculation of the co-occurrence degree of topics based on the binary matrix will be affected by the size effect of the topic probability itself. The size effect will make the topics that often appear in the document set have a high frequency of co-occurrence with other topics in reality. If the size effect is not eliminated, the final topic co-occurrence result will not reflect the real combination (i.e., co-occurrence) association degree between topics. The association strength standardization method used in the steps of the method is a measurement method for measuring the deviation of the observed co-occurrence frequency from the expected co-occurrence frequency under the assumption of independence. Therefore, the association strength standardization method can eliminate the size effect existing in the co-occurrence data.

[0151] The combination association matrix C recording the combination association degree between topics calculated by the association strength standardization method comprises the following steps:

[0152] S81: Set a threshold η, and set the element value less than η in the document-topic probability distribution matrix θ to 0, thereby filtering rare events, which is expressed by the formula:

[0153]

[0154] The reason and necessity of filtering rare events are explained as follows:

[0155] Since the association strength standardization method will deviate when facing rare events, in order to avoid the influence of rare events and obtain more accurate results, a threshold for filtering rare events is needed. The rare event refers to an event with a very small probability of occurrence.

[0156] An example of calculation bias of the association strength standardization method in the face of rare events: the number of times event A occurs is 1, and the number of times event B occurs is also 1, that is, event A and event B each occur only once. The case where event A and event B occur at the same time is exactly the case where event A and event B occur at the same time, which is expressed by the formula: A = B = AB = 1. Then, the probability of the occurrence of event A and event B together is calculated by the association strength standardization method as follows:

[0157]

[0158] Obviously, the probability value of the occurrence of event A and event B together calculated according to the association strength standardization method is a counterfactual, larger value.

[0159] An example: set the threshold η to 0.01, filter the rare events in the document-topic probability distribution matrix θ obtained in the S62 example based on η, and the filtered matrix θ is as follows:

[0160]

[0161] S82: Calculate the combination association matrix C of the combination association degree between the recorded topics by using the association strength standardization method:

[0162]

[0163] C is a symmetric matrix, and the calculation formula of the element value is:

[0164]

[0165] θ id is the element value in the document-topic probability distribution matrix θ, that is, the probability of the occurrence of topic i in a certain document d; θ jd is the element value in the document-topic probability distribution matrix θ, that is, the probability of the occurrence of topic j in a certain document d; ∑ d θ id θ jd is the probability of the occurrence of topic i and topic j together in the document set; ∑ d θ id is the total probability of the occurrence of topic i in the document set; ∑ d θ jd is the total probability of the occurrence of topic j in the document set; then, in order to ensure that there is no node self-connection in the network constructed based on the combination association matrix C subsequently, the diagonal element value of the combination association matrix C is processed to zero, that is:

[0166] cii = 0

[0167] Finally, the processed proximity association matrix C is obtained.

[0168] For example, based on the document-topic probability distribution matrix θ obtained in the example S81, the association strength standardization method is used to calculate and perform the 0 processing on the diagonal element values to obtain the combined association matrix C:

[0169]

[0170] S9: Based on the proximity association matrix P and the combined association matrix C, a dual network with two network levels is constructed and visualized; based on the multilinks vector in the dual network record the association degree between the topic pairs (any topic i and topic j in the K number of topics) a two-dimensional space is constructed; the threshold values λ p and λ c that divide the association degrees in the proximity network layer and the combined network layer are set, respectively; based on the threshold values λ p and λ c the two-dimensional space is divided into four different space regions, four association modes of the topic pairs are determined, and the four association modes include dual association, proximity dominant association, combined dominant association, and weak association.

[0171] The dual network is represented by the super adjacency matrix :

[0172]

[0173] wherein I is the unit matrix, that is:

[0174]

[0175] The unit matrix I in the non-diagonal position of the super adjacency matrix has no actual meaning and is only used to record and identify the one-to-one correspondence between the same nodes in the two network layers.

[0176] In the present application, the dual network The method steps of the application can realize embedding multiple technical association relationships into a network-based model containing different technologies and technical association relationships, so as to study the complex association between technologies. The existing network-based technical association analysis method uses a single-layer network model, and therefore is only suitable for exploring technical association analysis with only one type of association. However, the association between different technologies is complex and contains multiple types of association. The single-layer network model can only aggregate different types of association into one type of association for research, thereby losing the association information contained between different types of association. The double network can overcome this problem. The double network contains different network levels, and the nodes in different network levels are the same, but the connection relationship between the nodes is different. The nodes represent different subjects, and the connection relationship between the nodes represents the association relationship between the subjects. Therefore, the use of the double network can be used to represent two types of association between the same group of technologies, thereby providing a basis for analyzing the complex association mode between technologies.

[0177] In the application, the multi-link vector The method steps of constructing the two-dimensional space can be used to determine four types of association modes, thereby assisting in identifying potential technical opportunities. In the process of technological innovation, recombination of knowledge elements of an existing technical field, or reconfiguration of the combination mode of knowledge elements of a technical field that has been combined together, is an important way to give birth to new technical fields and new technical solutions. Therefore, exploring the association relationship and association mode between technologies plays an important role in the discovery of technical opportunities. The method steps can clarify the association mode between technical pairs (two different technologies) when there are two types of technical association by using a multi-link analysis method, thereby identifying potential technical opportunities.

[0178] Specifically, the double network is composed of two network levels, and the nodes in the two network levels are the same, and the nodes represent the K subjects obtained by S6. The links between the nodes in the two network levels respectively record the association degree corresponding to the different association relationships between the same group of nodes, i.e. the proximity association degree and the combination association degree between the K subjects 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 a one-to-one correspondence between the nodes representing the same node (i.e. the subject representing the technology) in the two network levels, and has no actual association degree significance.

[0179] For example, based on the proximity association matrix P obtained in the S7 example and the combination association matrix C obtained in the S8 example, a double network is constructed, which is represented by a super-adjacency matrix is represented as:

[0180]

[0181] Then, the dual network is visualized. Specifically, the constructed dual network is visualized based on the python toolkit pymnet. The constructed dual network is visualized as shown in Figure 3 .

[0182] Based on the multilink vector recording the degree of association between topic i and topic j in the dual network , a two-dimensional space is constructed. Specifically, the first element p in the multilink vector is set as the x-axis, and the second element c is set as the y-axis.

[0183] In the dual network , there is a multilink vector between any topic i and topic j, where p ij and c ij are the degree of association in the proximity network layer and the degree of association in the combined network layer of topic i and topic j, respectively, and p ij = p ji , c ij = c ji . Based on the two-dimensional features in the multilink vector , a two-dimensional space is constructed, where the first element p in the multilink vector is set as the x-axis, and the second element c is set as the y-axis.

[0185] For example, for the dual network constructed in the above example, the multilink vector of all topic pairs is:

[0186]

[0187] Based on the two-dimensional features in the multilink vector , a two-dimensional space is constructed. Specifically, the first element p in the multilink vector is set as the x-axis, and the second element c is set as the y-axis, and a two-dimensional space is constructed as shown in Figure 4 .

[0188] The threshold values λ p and λ c are set to divide the degree of association in the proximity network layer and the combined network layer into strong and weak, respectively. Specifically, the threshold value λ p in the proximity network layer can be set as the median, the third quantile, etc. of the degree of association in the proximity network layer; the threshold value λ c in the combined network layer can be set as the median, the third quantile, etc. of the degree of association in the combined network layer.c The threshold value λ c = 1. This is because the value calculated by the association strength standardization method in the calculation of the combination association matrix C has the following characteristics: the degree of co-occurrence of two subjects under the assumption of independence can be distinguished by whether the value is greater than 1. If the value is greater than 1, it means that the probability of co-occurrence of two subjects under the assumption of independence is higher than the expected value of the probability of co-occurrence of two subjects, i.e. the degree of association of co-occurrence (i.e. combination) is considered to be high, i.e. the degree of combination association is high; otherwise, the degree of association of co-occurrence (i.e. combination) is considered to be low, i.e. the degree of combination association is low.

[0189] For example, the threshold value λ p in the proximity network layer is set to the 80th percentile of the association degree value in the proximity network layer, i.e. the 80th percentile of the value of the element in the lower left corner (or upper right corner) of the proximity association matrix P. The threshold value λ c in the combination network layer is fixed at 1. The association degree thresholds in the proximity network layer and the combination network layer are represented by the following formulas, respectively:

[0190] λ p = 0.70, λ c = 1

[0191] Based on the set threshold values λ p and λ c , the two-dimensional space constructed by the multi-link vector is divided into four different space regions, thereby determining four association modes of the subject pair, including double association, proximity dominant association, combination dominant association and weak association. Note that the subjects represent different technologies, and the subject pair represents different technology pairs, so the association mode of the technology pair is the association mode between technologies.

[0192] The double association refers to the technology pair having more same or similar technical content and knowledge, and being frequently combined. Such technology pairs are likely to have been developed and utilized in a recombination manner, so the creative potential of these technology pairs is basically exhausted.

[0193] The proximity dominant association belongs to the technology pair that may have a substitution relationship, etc. For example, one technology is an updated iteration based on another technology, both of which solve similar or the same technical problems. However, in such a technology pair, low-cost and more reliable innovation activities can still be carried out in the form of combination.

[0194] The combination dominant association refers to the technology pair having low commonality or similarity of technical knowledge and elements, but still having recombination innovation activities and being more likely to bring about breakthrough innovation.

[0195] Weak connections refer to technology pairs that not only possess differentiated technical knowledge bases but are also rarely combined. Innovators can experiment with combining these technology pairs to achieve more novel technological innovations by integrating cross-disciplinary technical knowledge.

[0196] For example: Figure 5 As shown in the figure, the first quadrant (strong proximity association and strong combination association) is a dual association; the second quadrant (weak proximity association and strong combination association) is a combination-dominant association; the third quadrant (weak proximity association and weak combination association) is a weak association; and the fourth quadrant (strong proximity association and weak combination association) is a proximity-dominant association. Each sample point in the two-dimensional space is a multi-link vector corresponding to a technology pair (i.e., a topic pair). In the figure, they are identified by the label style "ij".

[0197] 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.

[0198] The above description is only a preferred embodiment of the present invention and does not constitute any form of limitation to the present invention. The scope of protection of the present invention shall be subject to the scope of protection of the claims. Although the present invention has been disclosed as above with preferred embodiments, it is not intended 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 shall still fall within the scope of the technical solution of the present invention.

Claims

1. A dual-technology correlation analysis method based on a dual network, characterized by: The following steps are involved: Extract technical keywords based on technical documents, build technical search formulas, and obtain technical keywords from technical databases based on technical search formulas. Original technical documents; right The text content in the original technical document is segmented and preprocessed for the first time to obtain preprocessed technical document sentences. Based on the preprocessed technical document sentences, a word embedding model is used for training to obtain word embeddings corresponding to the terms contained in the original technical document; Based on the The original technical documents are preprocessed for the second time 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 format, topic modeling is performed using the ETM model to obtain three results, including the topic-term probability distribution matrix , document-topic probability distribution matrix and the topic embedding matrix ; Based on the topic embedding matrix The cosine similarity calculation formula is used to measure the proximity between two topic embeddings, and the proximity correlation matrix that records the degree of proximity correlation between the topics is obtained. , based on the document-topic probability distribution matrix The combined association matrix of the combined association degree between the record topics is calculated by the association strength standardization method , where the association strength normalization method is a measurement method that measures the deviation between the observed co-occurrence frequency and the expected co-occurrence frequency under the independence assumption, based on the proximity association matrix and the combined incidence matrix , a dual network with two network layers is constructed , and for the dual network Visualization; based on dual network Multiple linkage vectors that record the degree of association between topic pairs Construct a two-dimensional space. and Themes and themes The degree of association in the proximity network layer and the degree of association in the combination network layer; setting the threshold for dividing the degree of association into strong and weak in the proximity network layer and the combination network layer respectively and ; Threshold-based and Divide the two-dimensional space into four different spatial regions and determine four association patterns of the subject pairs, the four association patterns including dual association, proximity-dominated association, combination-dominated association, and weak association; Dual Network By the hyperadjacency matrix express: ， in 。 2. The dual-technology correlation analysis method based on a dual network according to claim 1, characterized in that: The word embedding model is a Word2Vec model or a Glove model.

3. The dual-technology correlation analysis method based on a dual network according to claim 1, characterized in that: The specific steps of sentence segmentation and the first text preprocessing include: The Python toolkit NLTK is used to segment the text content of each original technical document to obtain technical document sentences with sentences as units; The python toolkit spaCy is used to perform the first text preprocessing on the technical document sentences to obtain the preprocessed technical document sentences.

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

5. The dual-technology correlation analysis method based on a dual network according to claim 1, 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 length of less than 3 characters, and non-noun terms in the original technical documents, restoring the remaining terms and converting them to lowercase, and deleting terms with a document frequency of less than 3 in the technical document collection. Finally, the preprocessed dictionary with a size of Technical documentation recorded in the form of bag of words.

6. The dual-technology correlation analysis method based on a dual network according to claim 1, characterized in that: Based on the word embedding and the technical documents in the bag-of-words format, the topic modeling process is performed through the ETM model to obtain three results, including the topic-term probability distribution matrix , document-topic probability distribution matrix and the topic embedding matrix The specific steps are as follows: The technical documents recorded in the word embedding and bag-of-words form are used as input data for training the ETM model, and the number of topics in the ETM model is determined by repeated experiments. The optimal setting value of this parameter; Based on the technical documents recorded in the word embedding and bag-of-words form as model input, the number of topics based on the ETM model parameter value Perform topic modeling and obtain output results in the form of 3 matrices: Topic-term probability distribution matrix : Row vector For the theme The probability distribution over all terms, where Theme Themes Document-topic probability distribution matrix : Row vector For Documents The probability distribution over all topics, where Theme documents; Topic Embedding Matrix : column vector For the theme The corresponding topic embedding of Theme Themes.

7. The dual-technology correlation analysis method based on a dual network according to claim 1, characterized in that: The cosine similarity calculation formula is used to measure the proximity between two topic embeddings, and the proximity correlation matrix that records the degree of proximity correlation between the topics is obtained. The specific steps are as follows: Use the cosine similarity calculation formula to obtain the proximity association matrix The element value in ; For a symmetrical square matrix, the calculation formula for the element value is: in, For the theme The corresponding theme embedding, For the theme The L-2 norm of the corresponding topic embedding, and For the theme With the theme The degree of proximity between them; By normalizing the proximity matrix The range of element values ​​in is scaled from [-1, 1] to [0, 1], and the normalization calculation formula is: is the normalized matrix element value; Proximity Correlation Matrix The diagonal element values ​​of are returned to 0, that is: Finally, the processed proximity correlation matrix is ​​obtained .

8. The dual-technology correlation analysis method based on a dual network according to claim 1, characterized in that: Based on the document-topic probability distribution matrix The combined association matrix of the combined association degree between the record topics is calculated by the association strength standardization method The specific steps are as follows: Set a threshold , the document-topic probability distribution matrix Smaller than The element value of is set to 0 to filter rare events. The formula is expressed as: The combined association matrix of the combined association degree between record topics is calculated using the association strength standardization method : For a symmetrical square matrix, the calculation formula for the element value is: is the document-topic probability distribution matrix The element value in the subject In the documentation The probability of occurrence in is the document-topic probability distribution matrix The element value in the subject In the documentation The probability of occurrence in For the theme With the theme The probability of co-occurrence in a document collection, For the theme The total probability of appearing in the document collection; For the theme The total probability of appearing in the document collection; Combination incidence matrix The diagonal element values ​​of are returned to 0, and the processed combined correlation matrix is ​​finally obtained. , the zero processing formula is: 。 9. The dual-technology correlation analysis method based on a dual network according to claim 1, characterized in that: Based on dual network Record topic and themes Steps to construct a two-dimensional space of multiple links based on the degree of association between them: In dual network For any topic and themes Construct multiple link vectors between , and Themes and themes The degree of association in the proximity network layer and the degree of association in the combination network layer, where , , ; Based on the multiple linkage vector The two dimensional features in construct a two-dimensional space, where The first element in As Axis, second element As axis.

10. The dual-technology correlation analysis method based on dual networks according to claim 1, characterized in that: The threshold Set to the median or third quartile or 80th percentile of the correlation value in the proximity network layer; the threshold Fixed to 1.

Images