Quantitative scientific research project approval screening method based on patent situation analysis

Abstract

The invention discloses a quantitative scientific research project screening method based on patent situation analysis, and the method comprises the steps: obtaining patent data, and carrying out the vectorization of a patent text through a pre-trained Chine-BERT model; the method comprises the following steps: performing dimension reduction processing on patent text feature vectors by using a UMAP algorithm, performing clustering analysis on the feature vectors after dimension reduction by using a K-Means + + algorithm, identifying professional technical themes and technical effect themes, merging documents of the same clustering category into a text unit, calculating the importance of each word by using a TF-IDF algorithm, and extracting the feature vectors of the patent text by using the TF-IDF algorithm. Extracting a theme keyword of each cluster; and a technology-effect matrix is constructed based on a clustering result, and a quantitative screening result of the scientific research project is obtained through calculation of a technology life cycle, technology innovation evaluation and market competitiveness evaluation. The objective of the invention is to provide a comprehensive and efficient scientific research project screening method through an advanced semantic understanding patent clustering method, and help to rapidly screen scientific research projects.

Description

Technical Field

[0001] The present invention relates to the field of technology evaluation technology, and in particular to a quantitative scientific research project screening method based on patent situation analysis. Background Art

[0002] At present, the research project justification is mainly carried out by combining peer review with application materials. This method belongs to the qualitative analysis based on expert experience and has corresponding disadvantages. First, the results of peer review often depend on the personal knowledge, experience, opinions and prejudices of the review experts, which may lead to the strong subjectivity of the evaluation results and the lack of unified standards and objectivity. Secondly, the fairness and rationality of the selection of reviewers are difficult to guarantee. Finally, scholars who already enjoy a high reputation and resources are more likely to gain recognition and support, while emerging scholars or scholars in marginal fields may face more difficulties and challenges. Summary of the invention

[0003] The purpose of the present invention is to provide a quantitative scientific research project screening method based on patent situation analysis. It aims to provide a comprehensive and efficient scientific research project screening method through advanced semantic understanding patent clustering method to help quickly screen scientific research project projects.

[0004] To achieve the above object, the present invention is implemented according to the following technical solutions:

[0005] The present invention comprises the following steps:

[0006] S1: Obtain patent data and use the pre-trained Chinese-BERT model to vectorize the patent text;

[0007] S2: Use the UMAP algorithm to reduce the dimension of the patent text feature vector, and use the K-Means++ algorithm to perform cluster analysis on the reduced feature vector to identify professional technical themes and technical efficacy themes;

[0008] S3: Remove stop words, merge documents of the same cluster category into one text unit, and use the TF-IDF algorithm to calculate the importance of each word, thereby extracting the topic keywords of each cluster;

[0009] S4: Construct a "technology-efficacy" matrix based on the clustering results, and obtain quantitative screening results of scientific research projects by calculating the technology life cycle, technology innovation evaluation and market competitiveness evaluation.

[0010] The beneficial effects of the present invention are:

[0011] The present invention is a quantitative scientific research project screening method based on patent situation analysis. Compared with the prior art, the present invention realizes quantitative scientific research project screening based on patent situation analysis through the following key technical points:

[0012] (1) The present invention aims to conduct comprehensive, objective and quantitative analysis of scientific research projects, avoiding the uncertainty caused by subjective judgment in current mainstream peer review methods. It is an auxiliary analysis method that takes into account the three levels of technology, market and technology development stage.

[0013] (2) The present invention uses advanced pre-trained models that have been pre-trained on large-scale corpora, which can better capture the deep semantic information in the text and significantly improve the accuracy of the semantic understanding of patent documents. In addition, after converting the patent abstract text into vectors, the present invention uses an efficient clustering algorithm to analyze these vectors, automatically identifying the professional and technical topics in the patent text, and providing users with a more accurate division of technical fields.

[0014] Through the above two techniques, the technology-effectiveness matrix is ​​obtained, and finally the innovation evaluation is calculated. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 is a flow chart for evaluating the technical innovation of the present invention;

[0016] Figure 2 It is a schematic diagram of the "technology-efficacy" matrix of the present invention;

[0017] Figure 3 It is a technology-efficiency matrix that uses patent data as the main display object;

[0018] Figure 4 This is a schematic diagram of the STC score ranking. DETAILED DESCRIPTION

[0019] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments. The illustrative embodiments and descriptions of the present invention are used to explain the present invention but are not intended to limit the present invention.

[0020] like Figure 1 As shown: The present invention is divided into three calculation models, Q = technology life cycle Q1 + technology innovation Q2 + market competitiveness Q3, and the calculation steps are as follows:

[0021] S1: Obtain patent data, and use the pre-trained Chinese-BERT model to vectorize the patent text; first, the present invention uses the Chinese-BERT pre-trained model optimized specifically for Chinese text to vectorize the text, and fine-tunes the pre-trained model at the same time. The specific steps are: select the cleaned patent text data, segment the text and create the input format of the mask language, and then use the Chinese-BERT model to perform field adaptation fine-tuning operations to adapt to the special language style and terminology of the patent document, so as to better understand the semantics of the patent text. In order to improve the fine-tuning effect, in this implementation case, about 10,000 patents related to "remote control and cluster control" are selected from a wide range. The parameter settings during the fine-tuning process include selecting AdamW as the optimizer, train_batch_size is 16, num_train_epochs is 3, warmup_steps is 500, learning_rate is 2e-5, weight_decay is 0.01, and mlm_probability is 0.15. Then, the fine-tuned Chinese-BERT model is used to convert the patent abstract and technical efficacy paragraphs into dense vector representations. These vectors can accurately capture the deep semantic features of the text. After the text is calculated by BERT, the text representation is obtained as follows: Where h is a vector in the vector set, M represents the number of words into which the text is segmented, and d h Represents the dimension of the latent vector h.

[0022] Among them, obtaining patent data includes:

[0023] 1. Select search tools and databases:

[0024] Select a suitable patent database according to your search objectives and needs. Different databases differ in terms of coverage, data update speed, and search functions, so you need to choose one based on your actual needs.

[0025] 2. Develop a search strategy:

[0026] Based on the analysis field and objectives, extract keywords that can reflect the subject content, including technical terms, product names, inventor names, etc. At the same time, consider using synonyms, related words, and extended keywords to increase the comprehensiveness of the search. Use logical operators (such as AND, OR, NOT) to combine keywords into search formulas to accurately express search requirements.

[0027] 3. Perform a search operation:

[0028] In the selected patent database, enter the constructed search formula. Based on the preliminary search results, adjust the keywords, search formula or select other search methods (such as classification search, applicant search, etc.) to obtain more accurate search results.

[0029] 4. Filter and organize search results:

[0030] The search results are deduplicated and relevant patent documents are screened out according to the analysis requirements. The screened patent documents are organized in a certain format (such as Excel table) to facilitate subsequent data analysis and processing.

[0031] S2: Use the UMAP algorithm to reduce the dimension of the patent text feature vector, and perform cluster analysis on the reduced feature vector through the K-Means++ algorithm to identify professional and technical topics and technical efficacy topics; the acquired text feature vector is then processed by the UMAP algorithm for dimensionality reduction to simplify the data structure and retain key information. In this implementation case, the parameter settings of the UMAP algorithm include n_neighbors as 15, n_components as 5, min_dist as 0.0, and metric as cosine. After the UMAP algorithm, the text vector can be reduced to 5 dimensions, which is helpful for subsequent clustering. Then use the K-Means++ algorithm to perform cluster analysis on the patent abstract and technical efficacy paragraphs respectively. K-Means++ is an improved K-Means algorithm that improves the clustering effect by optimizing the initialization process. The calculation steps of the algorithm are as follows:

[0032] A sample point is randomly selected from the data set as the first cluster center c1.

[0033] For each point x in the dataset i , calculate its shortest distance D(x i ), the calculation formula is:

[0034]

[0035] where x d and c jd They are point x i and cluster center c j The coordinate value in the nth dimension, N is the dimension of the data, and k is the number of set cluster centers.

[0036] The probability P(x) of selecting the next cluster center from a data point i ) and D(x i ) 2 The calculation formula is proportional to the square of:

[0037]

[0038] Where D(x i ) is the point x i The shortest distance to all selected cluster centers, ∑ x∈X D(x) is the sum of the distances from all data points to their nearest cluster center. This means that if a point is farther away from the nearest cluster center, then the probability of it being selected as the next cluster center is higher.

[0039] Repeat the above process until the maximum number of iterations is reached or the cluster center no longer changes.

[0040] In this implementation case, the number of cluster topics is set in the range of 10-20, which is determined after multiple experiments. Other parameters include init is k-means++, n_init is 10, max_iter is 300, and tol is 1e-4. Through this process, the present invention can identify the topic clustering results of the two dimensions of technology and efficacy from the patent abstract and the technical efficacy paragraph.

[0041] S3: Remove stop words, merge documents of the same cluster category into one text unit, and use the TF-IDF algorithm to calculate the importance of each word, thereby extracting the topic keywords of each cluster;

[0042] After removing stop words, documents of the same cluster category are merged into one text unit, and the importance of each word is calculated using the TF-IDF algorithm. This method can identify keywords that frequently appear in a specific cluster and are uncommon in other clusters, thereby extracting the topic keywords of each cluster. The specific calculation formula of TF-IDF is as follows:

[0043]

[0044] TF represents the frequency of a word in a document. The higher the frequency, the more times the word appears. IDF represents the inverse document frequency of a word. The fewer documents contain the word, the more the word can reflect the topic of the document. count(w) represents the number of occurrences of word w in document D. i The number of occurrences in |D i | indicates document D i The total number of words in the document collection, N represents the total number of documents in the document collection, Represents the number of documents that contain the term w among all documents.

[0045] 20 keywords are pre-selected for each cluster using the TF-IDF algorithm. In order to further refine the topic keywords, the present invention uses the MMR algorithm to optimize and select the keywords that best represent each topic. The MMR algorithm takes into account the relevance of the word to the topic and the distinction between different topics, ensuring that the selected keywords are both relevant and unique.

[0046] The calculation formula of the MMR algorithm is:

[0047] MMR=λ·Relevance(term)-(1-λ)·max[Diversity(term,term′)]

[0048] Where term is the candidate keyword currently being considered, term′ is a word in the selected keyword list, Relevance(term) is a function that evaluates the relevance of the candidate word to the topic, Diversity(term,term′) is a function that calculates the difference between the candidate word and the selected keyword, and λ is a parameter between 0 and 1 that determines the trade-off between relevance and diversity; the larger the λ value, the more the algorithm tends to select words with high relevance, and vice versa.

[0049] In the present invention, cosine similarity is selected as the Relevance (term) function, and the formula of cosine similarity is:

[0050]

[0051] Where A·B is the dot product of the two vectors A and B to be compared, and ∥A∥×∥B∥ is the L2 norm product of the two vectors, that is, the square root of the sum of the squares of all elements of each vector.

[0052] Select Jaccard distance as the Diversity(term,term′) function. The formula of Jaccard distance is:

[0053]

[0054] Where A and B are the two vectors A and B sets to be compared, |A∪B| is the number of elements in the intersection of A and B, and |A∩B| is the number of elements in the union of A and B.

[0055] After the MMR algorithm optimized the topic keywords, we finally obtained 10 keywords that can reflect the clustering theme.

[0056] S4: Based on the clustering results, a "technology-efficiency" matrix is ​​constructed, and the quantitative screening results of scientific research projects are obtained by calculating the technology life cycle, technology innovation evaluation and market competitiveness evaluation;

[0057] Domain experts perform operations such as word exclusion, word replacement, reordering, and manual summarization on the clustered keywords obtained above to obtain a professional expression of technical efficacy composed of keywords. In the implementation case, the final technical clustering themes include 16 categories such as "lifting equipment", "signal control", and "cluster management"; the efficacy clustering themes include 10 categories such as "appearance improvement", "improved portability", and "improved stability". The above clustering themes are used to obtain the technical points in this field, and a technical efficacy matrix is ​​constructed as follows: Figure 2 As shown, the clustered professional and technical themes are used as the horizontal axis of the matrix, the technical effectiveness themes are used as the vertical axis of the matrix, and the upper values ​​of the matrix are the number of co-occurring patents in the direction of technical effectiveness.

[0058] Technology Lifecycle Q1

[0059] The present invention calculates the technical growth rate (Q 11 ), technology maturity coefficient (Q 12 ), technical aging coefficient (Q 13 ) and the new technology characteristic coefficient (Q 14 ) The historical values ​​of the four indicators and the trend of data changes in the past three years can be used to determine the life cycle of the technology.

[0060] 1. Technology life cycle Q1 includes Q 11 , Q 12 , Q 13 , Q 14 Four dimensions.

[0061] Q 11 Technology growth rate indicator

[0062]

[0063] A F ——Number of invention patent applications in the current year

[0064] A p ——The sum of the number of invention patent applications in the technical field in the five years preceding the current year

[0065] A p =A t +A t-1 +A t-2 +A t-3 +A t-4 (t represents the current year)

[0066] Q 12 Technology maturity indicator

[0067]

[0068] A S ——Number of invention patent applications in the current year

[0069] Q 13 Technical aging indicators

[0070]

[0071] A W ——Number of design patent applications in the current year

[0072] Q 14 New technology characteristic indicators

[0073]

[0074] 2. Four calculations completed nearly four years of Q 11 , Q 12 , Q 13 , Q 14 Finally, the difference between the two adjacent years is used to obtain the change values ​​of the past three adjacent years.

[0075] Take t-3≤i≤t

[0076] ——Represents the Q of the i-th year 1X The value of (X = 1, 2, 3, 4)

[0077]

[0078] 3. According to Q 11 , Q 12 , Q 13 , Q 14 Change trend and get Q1 score.

[0079] ΔQ in the past three years 11 , ΔQ 12 , ΔQ 13 , ΔQ 14 Any three or more are always greater than or equal to

[0080] ΔQ in the past three years 11 , ΔQ 12 , ΔQ 13 , ΔQ 14 In the above example, any two values ​​are always greater than or equal to

[0081] ΔQ in the past three years 11 , ΔQ 12 , ΔQ 13 , ΔQ 14 In the above example, any one of them is always greater than or equal to

[0082] ΔQ in the past three years 11 , ΔQ12 , ΔQ 13 , ΔQ 14 In, all are less than or equal to

[0083] Evaluation of technological innovation: Considering that the innovation points of scientific research projects do not necessarily correspond to existing research directions, this paper uses the Chinese-BERT model to vectorize the text of scientific research project innovation points, professional and technical topics, and technical efficacy topics, and measures the similarity between project innovation points and technical points by calculating the cosine similarity between vectors. The topic of the technical point with the highest similarity to the innovation point is selected as the research subdivision direction of the project, and the innovation evaluation index Q2 of the scientific research project is quantitatively calculated. The intermediate score is calculated by interpolation method, and the calculation formula is:

[0084] Q2=P min +(xx min )*(P max -P min ) / (x max -x min )

[0085] Where x is the number of co-occurring patents with the most similar technology points to the project innovation point, x max and x min is the maximum and minimum number of co-occurring patents of the technology point theme in the matrix, P max and P min It is the maximum value (3 points) and the minimum value (0 point) of the innovation score.

[0086] If the subject similarity between the project's innovation points and all technical points is lower than 0.5, it indicates that the research direction of the scientific research project falls outside the technical effectiveness matrix. At this time, its research value and research risk have high uncertainty, so the technical innovation evaluation index Q2 is 0 points.

[0087] In view of the excellent applicability of the enterprise-level patent portfolio analysis method in the field of technological competitiveness assessment, this method can comprehensively and accurately reflect the technological competitiveness of the enterprise. Therefore, based on in-depth literature research and authoritative expert evaluation, this paper carefully selects appropriate measurement indicators from the two key dimensions of patent quality and patent activity to ensure a comprehensive assessment of the technological competitiveness of the enterprise. A ) and patent quality (P Q ) product to measure the static technological competitiveness (STC). In view of the core role of inventors in the patent R&D process, this invention takes the number of inventors as a quantitative indicator of patent R&D input when constructing a static technological competitiveness evaluation model, and takes the number of authorized patents as a quantitative standard of patent R&D output. Considering the balance and equal importance of R&D input and output, the weights of both in the model are set to 0.5.

[0088] 1. Position the research enterprise in the entire industry based on static competitiveness (STC).

[0089] STC=P A ×P Q (P A For patent activities, P Q Patent quality)

[0090] P A ——Patent activities

[0091] P Q ——Patent quality

[0092] P A =λ1×P1+λ2×P2+λ3×P3+λ4×P4+λ5×P5+λ6×P6

[0093] Among them, λ1, λ2, λ3, λ4, λ5, and λ6 are the weights of P1, P2, P3, P4, P5, and P6 respectively.

[0094] P1——Average application age, the average number of years from application to authorization for all patents

[0095] P2——Average survival age, the average number of years from application to the present

[0096] P3——Average technology scope, the average number of IPC classification numbers of all patents

[0097] P4——Average patent technology size, the average number of simple families of all patents

[0098] P5——Average citation frequency, the average number of citations of all patents

[0099] P6——Scientific relevance, the average number of non-patent literature cited by all patents

[0100] P Q =0.5×A1+0.5×A2

[0101] A1——Number of authorized patents (output activities), number of authorized patents

[0102] A2——Number of inventors (involved activities), the sum of the unique times that inventors appear

[0103] 2. According to the STC ranking, the ranking percentage of a certain enterprise among all enterprises is X, and the enterprise competitiveness index P is obtained based on X.

[0104]

[0105] 3. Based on the enterprise competitiveness index P and the technology life cycle index Q1, we can obtain the enterprise market competitiveness Q3.

[0106]

[0107] Filter ResultsQ

[0108] Q = α1×Q1+α2×Q2+α3×Q3 (α1, α2, α3 are the weights corresponding to Q1, Q2, Q3)

[0109]

[0110] Case Study:

[0111] In order to verify the feasibility of the scientific research project evaluation index system based on patent data, this paper takes the project of improving signal control stability in remote control and cluster control of a certain science and technology engineering company as an example to quantitatively demonstrate the feasibility of the project. The whole process of data acquisition and obtaining the scores in the evaluation system is explained, and the final conclusion is drawn.

[0112] 1. Data Acquisition

[0113] By collecting a large number of "remote control and cluster control" related patents from Zhihuiya, we can understand the characteristics of remote control and cluster control, and then conduct a detailed decomposition of remote control and cluster control. Through repeated retrieval and debugging, we determined the keywords and the final search formula. The keywords selected highly relevant terms in the field of remote control and cluster control, such as "remote", "cluster", "excavator" and other expressions. A total of 9,891 patent data from 2000 to 2023 were retrieved and downloaded from Zhihuiya as samples for this study.

[0114] The search formula is as follows:

[0115] TTL_ALL: (remote OR remote control OR cluster OR cluster) AND TAC_ALL: (excavator OR loader OR bulldozer OR scraper OR dumper OR roller OR compactor OR crane OR lift OR winch OR vibratory pile hammer OR hydraulic hammer OR pile driver OR diesel pile hammer OR drilling machine OR static pile driver OR paver OR spreader OR working machine OR working machine OR working machine OR construction machinery OR building machinery OR tower crane OR TBM OR pipe jacking machine OR rock drilling trolley OR wet spraying manipulator OR laying machine OR installation machine OR conveyor OR shield machine OR tunnel boring machine OR mixing station OR mixing station OR steel bar processing OR crane OR track laying machine OR bridge erection machine OR crusher OR excavator OR pile driver OR paver OR road roller OR loader OR transfer belt conveyor OR belt conveyor OR roofing machine OR excavator loader OR shovel milling machine OR dust removal trolley OR drilling trolley OR charging trolley OR maintenance trolley OR lining trolley OR laying vehicle OR working platform vehicle OR anchor trolley OR grouting equipment OR multi-function drilling rig OR spraying unit OR inverted bridge formwork trolley OR spraying vehicle OR slag transport vehicle)

[0116] 2. Remote control and cluster control project evaluation

[0117] For a project to improve the stability of signal control in remote control and cluster control of a certain science and technology engineering company, the technology life cycle evaluation index Q1, technology innovation evaluation index Q2, and enterprise market competitiveness evaluation index Q3 were calculated, and finally the scientific research project was evaluated.

[0118] 2.1 Technology Life Cycle Q1

[0119] Technology growth rate index Q in the past three years 11 The changes are shown in Table 1.

[0120] Table 1 Q in the past three years 11 change

[0121] years 2020 2021 2022 2023 <![CDATA[Q 11 ]]> 0.2507 0.2406 0.1987 0.1125 Difference between two consecutive years —— -0.0101 -0.0419 -0.0862

[0122] Technology maturity index Q in the past three years 12 The changes are shown in Table 2.

[0123] Table 2 Q in the past three years 12 change

[0124] years 2020 2021 2022 2023 <![CDATA[Q 12 ]]> 0.6779 0.5926 0.5873 0.5617 Difference between two consecutive years —— -0.0853 -0.0053 -0.0256

[0125] Technology aging index Q in the past three years 13 The changes are shown in Table 3.

[0126] Table 3 Q in the past three years 13 change

[0127] years 2020 2021 2022 2023 <![CDATA[Q 13 ]]> 0.9617 0.9589 0.9300 0.9255 Difference between two consecutive years —— -0.0028 -0.0289 -0.0045

[0128] New technology characteristic indicators Q in the past three years 14 The changes are shown in Table 4.

[0129] Table 4 Q in the past three years 14 change

[0130] years 2020 2021 2022 2023 <![CDATA[Q 14 ]]> 0.9636 0.9128 0.8866 0.8211 Difference between two consecutive years —— -0.0508 -0.0262 -0.0655

[0131] It can be seen from this that

[0132] 2.2 Technological Innovation Q2

[0133] A total of 9891 patent data from 2000 to 2023 were retrieved and downloaded from Zhihuiya as samples for cluster analysis. Due to the limitations of the system clustering model, some technologies and functions cannot be clustered. Finally, based on the valid data of 6190 "remote control and cluster control" related technology-function matrix, the following table shows the clustering model. Figure 3 As shown. The technical clusters include 16 categories such as "construction machinery", "lifting equipment", and "signal control"; the functional clusters include 10 categories such as "appearance improvement", "improved portability", and "improved stability". The project direction of this section is to improve the stability of signal control. Figure 3 It can be seen that it belongs to the third row "signal control" and the third column "improving stability". This innovation point is a blank point, so the Q2 score is 3 points.

[0134] Enterprise market competitiveness Q3

[0135] The project company is a certain technology engineering company. The company ranks as follows in the overall STC score ranking: Figure 4 shown.

[0136]

[0137]

[0138] 2.4 Screening Results Q

[0139]

[0140]

[0141] The technical solution of the present invention is not limited to the above-mentioned specific embodiments. All technical variations made according to the technical solution of the present invention fall within the protection scope of the present invention.

Claims

1. A quantitative scientific research project screening method based on patent situation analysis, characterized in that: The following steps are involved: S1: Obtain patent data and use the pre-trained Chinese-BERT model to vectorize the patent text; S2: Use the UMAP algorithm to reduce the dimension of the patent text feature vector, and use the K-Means++ algorithm to perform cluster analysis on the reduced feature vector to identify professional technical themes and technical efficacy themes; S3: Remove stop words, merge documents of the same cluster category into one text unit, and use the TF-IDF algorithm to calculate the importance of each word, thereby extracting the topic keywords of each cluster; S4: Based on the clustering results, a "technology-efficacy" matrix is ​​constructed, and the quantitative screening results of scientific research projects are obtained by calculating the technology life cycle, technology innovation evaluation and market competitiveness evaluation.

2. The quantitative scientific research project screening method based on patent situation analysis according to claim 1 is characterized by: The step S1 specifically includes: selecting the cleaned patent text data, segmenting the text and creating the input format of the mask language, and then using the Chinese-BERT model to perform domain adaptation fine-tuning operations, and then using the fine-tuned Chinese-BERT model to convert the patent abstract and technical efficacy paragraphs into dense vector representations. After the text is calculated by the Chinese-BERT model, the text representation is obtained as follows: Where h is a vector in the vector set, M represents the number of words into which the text is segmented, and d h Represents the dimension of the latent vector h.

3. The quantitative scientific research project screening method based on patent situation analysis according to claim 2 is characterized by: The step S2 is specifically as follows: S21: Randomly select a sample point from the data set as the first cluster center c1; For each point x in the dataset i , calculate its shortest distance D(x i ), the calculation formula is: where x d and c jd They are point x i and cluster center c j The coordinate value in the nth dimension, N is the dimension of the data, and k is the number of cluster centers set; S22: The probability P(x) of selecting the next cluster center from the data point i ) and D(x i ) 2 The calculation formula is proportional to the square of: Where D(x i ) is the point x i The shortest distance to all selected cluster centers, ∑ x∈X D(x) is the sum of the distances from all data points to their nearest cluster center; S23: Repeat the above steps S21 and S22 until the maximum number of iterations is reached or the cluster center no longer changes, thereby achieving the result of identifying thematic clustering results of the two dimensions of technology and efficacy from the patent abstract and the technical efficacy paragraph.

4. The quantitative scientific research project screening method based on patent situation analysis according to claim 3 is characterized by: The TF-IDF algorithm formula in step S3 is as follows: Where TF represents the frequency of a word in a document; IDF represents the inverse document frequency of a word, and count(w) represents the number of occurrences of word w in document D. i The number of occurrences in |D i | indicates document D i The total number of words in the document collection, N represents the total number of documents in the document collection, Represents the number of documents containing the term w among all documents; The TF-IDF algorithm is used to pre-select keywords for each cluster, and the MMR algorithm is used to optimize and select the keywords that best represent each topic; The calculation formula of the MMR algorithm is: MMR=λ·Relevance(term)-(1-λ)·max[Diversity(term,term ′ )] Where term is the candidate keyword currently being considered, term′ is a word in the selected keyword list, Relevance(term) is a function that evaluates the relevance of the candidate word to the topic, Diversity(term, term′) is a function that calculates the difference between the candidate word and the selected keyword, and λ is a parameter between 0 and 1; Select cosine similarity as the Relevance(term) function. The formula for cosine similarity is: Where A·B is the dot product of the two vectors A and B to be compared, and ∥A∥×∥B∥ is the L2 norm product of the two vectors, that is, the square root of the sum of all the elements of each vector; Select Jaccard distance as the Diversity(term,term′) function. The formula of Jaccard distance is: Where A and B are the two vectors A and B sets to be compared, |A∪B| is the number of elements in the intersection of A and B, and |A∩B| is the number of elements in the union of A and B; After the MMR algorithm optimizes the topic keywords, we finally get keywords that can reflect the clustering theme.

5. The quantitative scientific research project screening method based on patent situation analysis according to claim 4 is characterized by: Evaluation of technical innovation in step S4: The Chinese-BERT model is used to vectorize the text of the scientific research project innovation points, professional technical topics, and technical efficacy topics, and the similarity between the project innovation points and technical points is measured by calculating the cosine similarity between the vectors, and the topic of the technical point with the highest similarity to the innovation point is selected as the research subdivision direction of the project, and the scientific research project innovation evaluation index Q2 is quantitatively calculated, and the intermediate score is calculated by interpolation method, and the calculation formula is: Q2=P min +(xx min )*(P max -P min ) / (x max -x min ) Where x is the number of co-occurring patents with the most similar technology points to the project innovation point, x max and x min is the maximum and minimum number of co-occurring patents of the technology point theme in the matrix, P max and P min are the maximum and minimum values ​​of the innovation score.

Images