Risk technology identification method based on patent data

Abstract

The application discloses a kind of risk technology identification methods based on patent data.The present application is based on the definition of key core technology and risk technology, using the target industry field of patent data to be mined to construct IPC technology co-occurrence network, through K-core analysis algorithm and the statistics of node centrality, identify key core group, according to the characteristics of risk technology, pre-constructed index for measuring patent quality, through quality score statistics to identify risk technology group;And with the help of information extraction model, the patents in the risk technology group are extracted in fine-grained technical phrases, and the representative technical phrases of the final risk technology group are obtained by judging after text clustering.The method of the present application makes full use of the design indicators of each field of patent entry, and obtains fine-grained risk technology phrases by extracting the title and abstract text of the patent, which achieves good screening effect in the risk technology identification task and significantly reduces the cost of solving this task.

Description

Technical Field

[0001] This invention belongs to the fields of artificial intelligence, data mining, deep learning, and risk technology identification, and particularly relates to a risk technology identification method based on patent data. Background Technology

[0002] Key core technologies refer to strategic high technologies that, in the context of international competition, have a significant impact on a specific country's economic, technological, and national security advantages. These are high-tech fields where a particular country engages in fierce competition with its rivals for dominance and long-term competitive advantage, generally characterized by strong strategic and pioneering nature, cutting-edge technology, and high value. Accurately identifying risky technologies in a specific country's industrial sectors helps to rationally allocate innovation resources, break through technological blockades by competitors, and enhance a country's overall competitiveness, thus possessing significant social value.

[0003] Currently, research on risk technology identification mainly includes three aspects: qualitative research based on expert analysis, quantitative research based on indicator statistics or network analysis, and research combining qualitative and quantitative methods.

[0004] Qualitative analysis primarily relies on expert knowledge and experience to make judgments, employing methods such as questionnaires and the Delphi method to collect expert opinions and identify risk technologies in specific fields. This type of method has been used early in risk technology identification research, but because it depends entirely on experts and scholars in specific fields, its practical applicability is relatively weak.

[0005] Quantitative analysis mainly combines data such as invention patents and scientific research papers, and uses techniques such as indicator system methods, text mining, and complex networks to analyze and identify risky technologies in specific fields.

[0006] (1) Indicator system: Starting from various types of technology definitions, the main characteristics of the technology are clarified, and then various objective or subjective indicators are used to represent various characteristics in order to achieve the purpose of technology identification.

[0007] (2) Complex network analysis: Complex networks are constructed using co-citation relationships, citation coupling relationships, direct citation relationships, and co-occurrence relationships of IPC classification numbers in scientific and technological literature and invention patents. Then, cluster analysis, link prediction, K-core analysis, node network feature analysis and other methods are used for technology identification.

[0008] (3) Text mining: Research is conducted using scientific and technological texts to extract keywords, themes, semantic structures, etc. to identify technologies that meet certain characteristics. The main research methods include keyword statistical analysis, co-word analysis, emerging word monitoring, theme modeling, SAO structure analysis, and technology form analysis.

[0009] Currently, there is no method that can combine objective indicators, complex network analysis, and text mining techniques for risk identification. Therefore, the identification performance is low and the granularity is coarse, requiring further improvement. This invention focuses on combining objective indicators and network analysis to screen the scope of patented technologies, and using text mining to extract and cluster fine-grained technical phrases to identify key core risk technologies within the field. Summary of the Invention

[0010] The purpose of this invention is to solve the problems existing in the prior art and to provide a risk technology identification method based on patent data.

[0011] To achieve the above-mentioned objectives, the present invention specifically adopts the following technical solution:

[0012] A method for identifying risky technologies based on patent data includes the following steps:

[0013] S1. Obtain technical keywords within the target industry sector. After translating and performing synonym substitution on the technical keywords, search the global patent database for patents whose titles or abstracts contain the technical keywords. Obtain the patent classification number within the target industry sector according to the technology classification system of the target industry sector. Obtain the corresponding patents according to the patent classification number. Combine the patents retrieved through technical keywords and the patents obtained according to the patent classification number as the patent data to be mined in the target industry sector. The target industry sector is the industry sector that needs to be mined for technology risk.

[0014] S2. Construct an IPC technology co-occurrence network using the patent data to be mined obtained in S1. The IPC technology co-occurrence network is a graph network consisting of nodes, edges, and edge weights. The nodes are IPC technology groups that have appeared in the patent classification numbers of the patent data to be mined. The edges represent the co-occurrence relationship between two IPC technology groups. The weight of the edge is the number of times the two IPC technology groups co-occur in the patent classification numbers of the patent data to be mined.

[0015] S3. Perform node centrality statistics on each node in the IPC technology co-occurrence network, and construct the first original node set using the nodes with the highest node centrality ranking in the IPC technology co-occurrence network; perform K-core analysis on the IPC technology co-occurrence network to obtain the maximum K-core subgraph, and construct the second original node set using the nodes in the maximum K-core subgraph; take the intersection of the first original node set and the second original node set to obtain the key core node set, and take the IPC technology groups in the key core node set as the key core groups; recall all patents in the patent data to be mined that use the key core groups as the main IPC classification numbers, and take them as the key core patent data to be mined;

[0016] S4. After pre-constructing indicators to measure patent quality based on the characteristics of risky technologies, calculate the quality score of all key core patent data. In each key core group, the core patent data with the highest quality score are regarded as high-quality patents for each key core group. The key core groups that meet the preset requirements of the evaluation indicators are regarded as risk technology groups to be explored. The evaluation indicators are the patent market share, quality score ratio and national gap of high-quality patent market share within the key core group.

[0017] S5. After pre-designing the schema constraints, use the fine-tuned information extraction model to extract fine-grained technical phrases from the patent texts within the risk technology group. Then, cluster the extracted fine-grained technical phrases into a set of technical phrase clusters, taking the risk technology group as the unit. Calculate and rank the quality scores of each technical phrase cluster in the set. Select the technical phrase clusters with the highest quality scores in each risk technology group as representative technical phrases for experts to analyze and verify the risk technologies in the patent data.

[0018] Based on the above scheme, each step can be implemented in the following preferred manner.

[0019] As a preferred option, in step S1, the technology classification system of the target industry field refers to an officially issued document that includes the reference relationship between the technology field and the international patent classification system.

[0020] As a preferred option, the specific process of step S3 is as follows:

[0021] S31. The centrality of a node is the sum of the weights of its neighboring edges, calculated using the following formula:

[0022] Deg(v)=Σ {v,u}∈E w(v,u)

[0023] Where Deg(v) is the sum of the weights of the neighboring edges of node v in the point set V of the IPC technology co-occurrence network; u represents a node in the point set V of the IPC technology co-occurrence network that is connected to node v; E is the edge set of the IPC technology co-occurrence network; w(v,u) represents the weight of the edge formed by nodes v and u; {v,u}∈E represents an edge in the edge set E of the IPC technology co-occurrence network formed by nodes v and u;

[0024] S32. In the IPC technology co-occurrence network, select the nodes with the highest node centrality ranking to construct the first original node set. The calculation process of the first original node set DegSet is as follows:

[0025] c=min(|V|*b,|{v∈V|Deg(v)>a}|)

[0026]

[0027] Where |V| represents the number of nodes in the point set V of the IPC technology co-occurrence network; b is a set threshold for the percentage of nodes; a is a set threshold for the degree of nodes; argmax c This represents the top c nodes with the highest node centrality values ​​in the point set V of the IPC technology co-occurrence network.

[0028] S33. In the k-th analysis of the K-core analysis, remove all nodes in the IPC co-occurrence network whose sum of neighboring edge weights is less than k and their connected edges to obtain the original subgraph. Recalculate the neighboring edge weights of the remaining nodes in the original subgraph until there are no nodes in the original subgraph whose neighboring edge weights are less than k, thus obtaining the subgraph of the k-th analysis; where the node set V of the subgraph of the k-th analysis is... k for:

[0029]

[0030] In the formula, G represents the sum of the weights of the neighboring edges of node u in the subgraph analyzed in the k-th iteration; k This represents the subgraph used in the k-th analysis.

[0031] S34. Continue the analysis until the k-th... max The k-th network in the +1st analysis shows the co-occurrence of IPC technology. max When the subgraph of the +1 analysis is empty, k max As the maximum number of times K-core analysis is performed, the kth... max The subgraph analyzed in this step is the largest K-kernel subgraph of the IPC technology co-occurrence network;

[0032] S35. Take the set of nodes in the maximum K-kernel subgraph as the second original node set, and combine the first original node set DegSet with the second original node set. Find the intersection to obtain the key core node set NodeSet;

[0033] S36. Recall all patents in the patent data to be mined that use the IPC technical group in the set of key core nodes as the IPC main classification number, and use them as key core patent data to be mined.

[0034] As a preferred option, the specific process of step S4 is as follows:

[0035] S41. Using the entropy method, weights are assigned to each selected patent quality indicator, and the quality score of all key core patent data is calculated. If there are n key core patent data items and m patent quality indicators, then the quality score S of the i-th key core patent data item is... i The calculation process is as follows:

[0036]

[0037] Among them, W j The weight of the indicator representing the quality of the j-th patent; The standardized value of the indicator representing the quality of the i-th key core patent data and the j-th patent.

[0038] S42. Taking key core groups as units, a pre-set percentage threshold p is set, and the patents with the highest quality scores in each key core group are regarded as high-quality patents of the key core group, and the evaluation indicators of each key core group are calculated.

[0039] The formula for calculating the country-specific differences in patent market share within key core groups is as follows:

[0040]

[0041] Among them, M i′ This represents the gap in patent market share between the target country within the key core group (classification number i′) and the country with the highest number of patent applications; X i′,target This indicates the number of patents held by the target country within the key core group of classification number i′; X i′,top1 The number of patents held by the top-ranked patent applicant country within the key core group of classification number i′; COUN represents the total number of patent applicant countries within the key core group of classification number i′; X i′,coun This represents the number of patents held by the country holding the first patent application in the key core group with classification number i′.

[0042] The formula for calculating the country-specific differences in the percentage of quality scores within key core groups is as follows:

[0043]

[0044] Among them, C i′ This indicates the difference in patent quality score percentage between the target country within the key core group (classification number i′) and the top-ranked patent applicant country; PAT target PAT top1 PAT i′ S represents the target country, the top-ranked patent applicant country, and the total number of patents from all patent applicant countries within the key core group of classification number i′, respectively; pat,target S pat,top1 S pat These represent the quality scores of the key core patent data for the target country, the top-ranked patent applicant country, and all patent applicant countries, respectively.

[0045] The formula for calculating the country-specific differences in market share of high-quality patents within key core groups is as follows:

[0046]

[0047] Among them, L i′ This represents the gap in market share of high-quality patents between the target country within the key core group (classification number i′) and the country ranked first in patent applications; Y i′,target Y represents the number of high-quality patents in the target country within the key core group of classification number i′; i′,top1 Y represents the number of high-quality patents held by the top-ranked patent application country within the key core group of classification number i′; i′,coun This represents the number of high-quality patents in the key core group with classification number i′ of the country that filed the patent application;

[0048] S44. Determine whether the three evaluation indicators of each key core group simultaneously meet the set thresholds corresponding to each evaluation indicator: if they simultaneously meet the thresholds, then the key core group is identified as a risk technology group to be explored.

[0049] Preferably, in step S4, the weight W of the index for the quality of the j-th patent is... j The calculation method is as follows:

[0050]

[0051] Among them, Q j The entropy value, representing the quality index of the j-th patent, is calculated as follows:

[0052]

[0053] Here, -1 represents the reciprocal.

[0054] Preferably, in step S4, the standardized value of the indicator of the quality of the j-th patent in the i-th key core patent data is used. The calculation method is as follows:

[0055]

[0056] Wherein, min(x) j ),max(x j ) represent the minimum and maximum values ​​of the quality index of the j-th patent among all key core patent data.

[0057] As a preferred option, in step S4, a total of 10 indicators for measuring patent quality are selected, namely, the number of patent families, the number of countries where patent families are deployed, the number of countries where patents are cited, the average annual citation frequency, the number of citations, the coverage of IPCs, the number of inventors, the number of countries of invention, the number of citations, and the number of claims.

[0058] Preferably, after obtaining the risk technology groups to be explored in step S4, the entropy method is used to assign weights to the strength indicators of each selected patent application agency, calculate the strength score of each patent application agency in each risk technology group, and select the patent application agency with the highest strength score in each risk technology group as the leading agency to verify the effectiveness of identifying risk technologies using patent quality indicators in step S4; wherein, the strength indicators of the patent application agency are the total number of patents, the sum of patent quality scores, and the total number of high-quality patents of the patent application agencies in each risk technology group;

[0059] Apps, the leading institution within the vth risk technology group v The methods for obtaining it are as follows:

[0060]

[0061] Where N_APPS is the set threshold for the number of leading institutions; argmax N_APPS The top N_APPS patent application agencies have the largest values; A(app) represents the strength score of the app-th patent application agency; Apps represents the number of patent application agencies in the v-th risk technology group.

[0062] As a preferred option, the specific process of step S5 is as follows:

[0063] S51. Pre-acquire a batch of labeled patent data. The entity extraction of the pre-trained information extraction model M(θ) is fine-tuned using labeled patent data; where x txt For the original text, y txt This is a list of technical phrases that have been annotated; TXT represents the total number of annotated texts.

[0064] S52. Extract patent texts from the risk technology group to be explored. And the schema constraints are input into the information extraction model M(θ) after entity extraction fine-tuning. * Extract fine-grained technical phrases from the patent text.

[0065] S53. Fine-grained technical phrases extracted from the same patent text are processed according to the information extraction model M(θ). * The given probability sorting process is as follows:

[0066]

[0067]

[0068] in, These represent the 1st, 2nd, ..., nsth fine-grained technical phrases, respectively. Let represent the probabilities of the 1st, 2nd, ..., nsth fine-grained technical phrases given the patent text and information extraction model, respectively.

[0069] S54. Select the top 10 fine-grained technical phrases with the highest probability from the sorted fine-grained technical phrases as the patent text for the risk technology group. List of technical phrases Wherein, Phrase represents the preset threshold for the number of fine-grained technical phrases;

[0070] S55. After obtaining the corresponding list of technical phrases for each patent text in each risk technology group, hierarchical clustering is performed on the fine-grained technical phrases in all technical phrase lists within each risk technology group, and the fine-grained technical phrases in all technical phrase lists within each risk technology group are converted into text vectors. The cosine similarity between any two text vectors is calculated, and the distance metric between the two text vectors is obtained after converting the cosine similarity calculation result. Clustering is stopped when the minimum distance between the text vectors of any two clusters exceeds the preset text vector threshold, and a set of technical phrase clusters is obtained.

[0071] The set of technical phrase clusters C within the rg-th risk technology group rg Represented as:

[0072]

[0073] Among them, C r C g ∈C rg Representing the set of technical phrase clusters C rg The text contains two distinct technical phrase clusters; a1, b1, a2, b2 represent arbitrary fine-grained technical phrases within the corresponding technical phrase clusters; d(a1, b1) and d(a2, b2) both represent the distance metric between two text vectors; β represents a preset text vector threshold; ∧ represents simultaneous satisfaction; r and g represent the set of technical phrase clusters C. rg Indexes of two different technical phrase clusters;

[0074] S56. Calculate the score of each technical phrase cluster within each risk technical group, and select the technical phrase cluster with the highest score as the representative technical phrase of the risk technical group.

[0075] The rg-th risk technology group, the r-th technology phrase cluster C r Score The calculation formula is as follows:

[0076]

[0077] Where NUM represents the total number of patents in the rg-th risk technology group; T t S represents the list of technical phrases in the t-th patent within the rg-th risk technology group; t This represents the quality score of the t-th patent within the rg-th risk technology group; ∩ represents the empty set; ∩ represents the intersection of the sets.

[0078] The representative technical phrase of the rgth risk technology group is RepTech. rg as follows:

[0079]

[0080] Where Kk is the set threshold for the number of phrases, and argmax Kk This represents the top Kk technical phrase clusters with the highest scores.

[0081] Compared with the prior art, the present invention has the following advantages:

[0082] Currently, there is no method that simultaneously combines objective indicators, complex network analysis, and text mining techniques for risk technology identification. Inspired by existing risk technology identification work, this invention constructs an IPC (Independent Patent Case) technology co-occurrence network using domain patent data. Network analysis is used to filter IPC groups, and the objective attributes of patents are leveraged to calculate a comprehensive patent score using entropy. IPC groups containing technological risks are selected based on both quantity and quality differences. An information extraction model is then used to perform fine-grained technical phrase extraction and text clustering on patents within these risk groups. This invention provides a method for automatically identifying risk technologies in a specific domain based on patent data. This invention reduces the time and cost spent on risk technology screening, narrows the scope of domain-specific risk technology identification, provides a reliable preliminary candidate set for expert verification, and improves identification efficiency and accuracy. Attached Figure Description

[0083] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0084] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below. Technical features in the various embodiments of the present invention can be combined accordingly without mutual conflict.

[0085] In the description of this invention, it should be understood that the terms "first" and "second" are used only for descriptive purposes and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined with "first" and "second" may explicitly or implicitly include at least one of those features.

[0086] like Figure 1 As shown, in a preferred embodiment of the present invention, the risk technology identification method based on patent data includes the following steps S1 to S5. The specific implementation process of each step will be described in detail below.

[0087] 1. Domain IPC mapping and patent data acquisition

[0088] S1. Obtain technical keywords within the target industry sector. After translating and performing synonym substitution on the technical keywords, search the global patent database for patents whose titles or abstracts contain the technical keywords. Obtain the patent classification number within the target industry sector according to the technical classification system of the target industry sector. Obtain the corresponding patents according to the patent classification number. Combine the patents retrieved through technical keywords and the patents obtained according to the patent classification number as the patent data to be mined in the target industry sector. The target industry sector is the industry sector that needs to be mined for technical risks.

[0089] It should be noted that in step S1 of this invention, the technology classification system of the target industry field refers to the officially issued document that includes the reference relationship between the technology field and the international patent classification system. In this embodiment, the patent classification number in the target industry field can be obtained through the "Key Digital Technology Patent Classification System (2023)" and the "Reference Relationship Table between Strategic Emerging Industries Classification and International Patent Classification (2021)" issued by the State Intellectual Property Office.

[0090] In this embodiment, the risk technology type of concern is the integrated circuit field. Therefore, it is necessary to collect patent data in the integrated circuit field in advance to construct a co-occurrence network for risk identification. First, based on literature research, keyword lists for each sub-field are compiled around the materials, design, manufacturing, packaging, and testing stages of the integrated circuit industry, and patent titles and abstracts are retrieved. Second, referring to the technology patent classification system table in the "Key Digital Technology Patent Classification System (2023)" document, patents in the international patent classification numbers involved in the integrated circuit industry field are retrieved.

[0091] 2. Construction of Patent IPC Co-occurrence Network

[0092] S2. Construct an IPC technology co-occurrence network using the patent data to be mined obtained in S1. The IPC technology co-occurrence network is a graph network consisting of nodes, edges, and edge weights. The nodes are IPC technology groups that have appeared in the patent classification numbers of the patent data to be mined. The edges represent the co-occurrence relationship between two IPC technology groups. The weight of the edge is the number of times the two IPC technology groups co-occur in the patent classification numbers of the patent data to be mined.

[0093] It should be noted that in step S2 of this invention, the IPC technology group refers to the fifth-level subcategory (i.e., the group of classification numbers) in the International Patent Classification (IPC). A co-occurrence relationship between two IPC technology groups means that a patent document appears in both IPC technology groups simultaneously. When constructing the IPC technology co-occurrence network, the citation frequency of the patent data to be mined affects the risk technology identification effect; therefore, this embodiment only uses patent data with a citation frequency greater than 0. The nodes in the IPC technology co-occurrence network are the IPC technology groups that have appeared in the IPC classification numbers of this batch of patents. The edges in the diagram represent the co-occurrence of two technology groups in the classification number of a patent, and the weight of the edge is the total number of times these two technology groups co-occur in the same patent classification number in this batch of patents, as shown in Table 3. Based on the patent classification number of the patent data to be mined, whenever two IPC technology groups appear simultaneously in the IPC classification number of the same patent, the weight of the edge between the nodes of these two IPC technology groups is increased by 1.

[0094] Table 3. Examples of Networks Where IPC Technologies Co-occur

[0095]

[0096] 3. Analysis of IPC technology co-occurrence network indicators and node selection

[0097] S3. Perform node centrality statistics on each node in the IPC technology co-occurrence network, and construct the first original node set using the nodes with the highest node centrality ranking in the IPC technology co-occurrence network; perform K-core analysis on the IPC technology co-occurrence network to obtain the maximum K-core subgraph, and construct the second original node set using the nodes in the maximum K-core subgraph; take the intersection of the first original node set and the second original node set to obtain the key core node set, and take the IPC technology groups in the key core node set as the key core groups; recall all patents in the patent data to be mined that use the key core groups as the main IPC classification numbers, and take them as the key core patent data to be mined.

[0098] It should be noted that step S3 of this invention consists of three sub-steps: node centrality calculation, K-core analysis, and patent recall. The specific process is as follows:

[0099] S31. Node Centrality Calculation: Node centrality is the sum of the weights of all the neighboring edges of a node. The calculation formula is as follows:

[0100] Deg(v)=∑ {v,u}∈E w(v,u)

[0101] Where Deg(v) is the sum of the weights of the neighboring edges of node v in the point set V of the IPC technology co-occurrence network, i.e., the node centrality; u represents a node connected to node v in the point set V of the IPC technology co-occurrence network; E is the edge set of the IPC technology co-occurrence network; w(v,u) represents the weight of the edge formed by nodes v and u; {v,u}∈E represents an edge formed by nodes v and u in the edge set E of the IPC technology co-occurrence network.

[0102] In this embodiment, the node centrality calculation step counts the centrality of nodes in the IPC technology co-occurrence network constructed in the previous step, and selects the nodes with the highest centrality ranking as candidates for the key core group.

[0103] S32. In the IPC technology co-occurrence network, select the nodes with the highest node centrality ranking to construct the first original node set. The calculation process of the first original node set DegSet is as follows:

[0104] c=min(|V|*b,|{v∈V|Deg(v)>a}|)

[0105]

[0106] Where |V| represents the number of nodes in the point set V of the IPC technology co-occurrence network; b is a set threshold for the percentage of nodes; a is a set threshold for the degree of nodes; argmax c This represents the top c nodes with the highest node centrality values ​​in the point set V of the IPC technology co-occurrence network.

[0107] In this invention, the threshold range for the percentage of nodes is related to the structural size of the graph network, and the threshold for the degree of nodes is related to the structural size of the graph network, but is not limited.

[0108] S33, K-core Analysis: In the k-th analysis of the K-core analysis, all nodes in the IPC co-occurrence network whose sum of neighboring edge weights is less than k and their connected edges are removed to obtain the original subgraph. The neighboring edge weights of the remaining nodes in the original subgraph are recalculated until there are no nodes in the original subgraph with neighboring edge weights less than k, resulting in the subgraph of the k-th analysis. The node set V of the subgraph in the k-th analysis is... k for:

[0109]

[0110] In the formula, G represents the sum of the weights of the neighboring edges of node u in the subgraph analyzed in the k-th iteration; k This represents the subgraph of the k-th analysis.

[0111] S34. Continue the analysis until the k-th... max The k-th network in the +1st analysis shows the co-occurrence of IPC technology. max When the subgraph of the +1 analysis is empty, k max As the maximum number of times K-core analysis is performed, the kth... max The subgraph analyzed is used as the largest K-core subgraph of the IPC technology co-occurrence network.

[0112] In this embodiment, starting from k=1, all nodes in the IPC technology co-occurrence network whose sum of neighboring edge weights is less than k, along with their connected edges, are removed. For each node removed, the neighboring edge weights of the remaining nodes are recalculated. This process is repeated until no node in the IPC technology co-occurrence network has a neighboring edge weight less than k, resulting in the subgraph analyzed in the kth iteration. Let k=k+1, and continue until the kth iteration... max The K-core subgraph of the IPC co-occurrence network is empty at +1 analysis time. max As the maximum number of K-core analyses, the largest K-core subgraph of the IPC technology co-occurrence network is finally obtained, whose node set satisfies:

[0113]

[0114] in, It is an empty set; k max This represents the largest integer in the IPC co-occurrence network that contains a K-core subgraph, i.e., the maximum number of K-core analyses.

[0115] S35. Patent Recall: Using the set of nodes in the maximum K-kernel subgraph as the second original node set, and combining the first original node set DegSet with the second original node set. Taking the intersection, we obtain the key core node set NodeSet, represented as follows:

[0116]

[0117] S36. Recall all patents in the patent data to be mined that use the IPC technical group in the set of key core nodes as the IPC main classification number, and use them as key core patent data to be mined.

[0118] 4. Patent Score Calculation and Country-Specific Differences

[0119] S4. After pre-constructing indicators to measure patent quality based on the characteristics of risky technologies, calculate the quality score of all key core patent data. In each key core group, the core patent data with the highest quality score are regarded as high-quality patents corresponding to each key core group. The key core groups that meet the preset requirements of the evaluation indicators are regarded as risk technology groups to be explored. The evaluation indicators are the national differences in patent market share, quality score ratio, and high-quality patent market share within the key core group.

[0120] It should be noted that in step S4 of this invention, the characteristics of the risky technology include, but are not limited to, strategic pioneering nature, cutting-edge technology, and complex innovation. These characteristics can be evaluated through the following patent attributes: number of patent families, number of countries where patent families are deployed, number of cited countries, average annual citation frequency, number of citations, number of inventors, number of invention countries, number of citations, number of claims, and IPC coverage. In this embodiment, based on the strategic pioneering nature that the risky technology should meet, three indicators are selected: number of patent families, number of countries where patent families are deployed, and number of cited countries. Based on the cutting-edge technology that the risky technology should meet, three indicators are selected: average annual citation frequency, number of citations, and IPC coverage. Based on the complex innovation that the risky technology should meet, four indicators are selected: number of inventors, number of invention countries, number of citations, and number of claims, for a total of 10 objective patent indicators.

[0121] It should be noted that step S4 of this invention consists of three sub-steps: patent quality score calculation, country-specific gap comparison, and verification by leading institutions. The specific process is as follows:

[0122] S41. Patent Quality Score Calculation: The entropy method is used to assign weights to the selected patent quality indicators, and the quality score of all key core patent data is calculated. If there are n key core patent data items and m patent quality indicators, then the quality score S of the i-th key core patent data item is... i The calculation process is as follows:

[0123]

[0124] Among them, W j The weight of the indicator representing the quality of the j-th patent; The standardized value of the quality index of the j-th patent, representing the i-th key core patent data.

[0125] In this invention, the weight W of the quality index of the j-th patent is... j The calculation method is as follows:

[0126]

[0127] Among them, Q jThe entropy value, representing the quality index of the j-th patent, is calculated as follows:

[0128]

[0129] Here, -1 represents the reciprocal.

[0130] In this invention, the standardized value of the indicator of the quality of the i-th key core patent data and the j-th patent is... The calculation method is as follows:

[0131]

[0132] Wherein, min(x) j ),max(x j ) represent the minimum and maximum values ​​of the quality index of the j-th patent among all key core patent data.

[0133] S42. Country-specific gap comparison: Taking key core groups as units, a pre-set percentage threshold p is set. The patents with the highest quality scores in each key core group are considered as high-quality patents of the key core group, and the evaluation indicators for each key core group are calculated.

[0134] The formula for calculating the country-specific differences in patent market share within key core groups is as follows:

[0135]

[0136] Among them, M i， This represents the gap in patent market share between the target country within the key core group (classification number i′) and the country with the highest number of patent applications; X i′,target This indicates the number of patents held by the target country within the key core group of classification number i′; X i′,top1 The number of patents held by the top-ranked patent applicant country within the key core group of classification number i′; COUN represents the total number of patent applicant countries within the key core group of classification number i′; X i′,coun This represents the number of patents held by country number coun in the key core group of classification number i′.

[0137] The formula for calculating the country-specific differences in the percentage of quality scores within key core groups is as follows:

[0138]

[0139] Among them, C i′ This indicates the difference in patent quality score percentage between the target country within the key core group (classification number i′) and the top-ranked patent applicant country; PAT target PAT top1 PAT i′S represents the target country, the top-ranked patent applicant country, and the total number of patents from all patent applicant countries within the key core group of classification number i′, respectively; pat,target S pat,top1 S pat These represent the quality scores of the key core patent data for the target country, the top-ranked patent applicant country, and all patent applicant countries, respectively.

[0140] The formula for calculating the country-specific differences in market share of high-quality patents within key core groups is as follows:

[0141]

[0142] Among them, L i′ This represents the gap in market share of high-quality patents between the target country within the key core group (classification number i′) and the country ranked first in patent applications; Y i′,target Y represents the number of high-quality patents in the target country within the key core group of classification number i′; i′,top1 Y represents the number of high-quality patents held by the top-ranked patent application country within the key core group of classification number i′; i′,coun This represents the number of high-quality patents held by country coun in the key core group of classification number i′.

[0143] S44. Determine whether the three evaluation indicators of each key core group simultaneously meet the set thresholds corresponding to each evaluation indicator: if they simultaneously meet the thresholds, then the key core group is identified as a risk technology group to be explored.

[0144] The RiskGroup, which consists of risk technology teams, is represented as follows:

[0145] RiskGroup = {v∈NodeSet|M i′ ≤α M ∧C i′ ≤α C ∧L i′ ≤α L}

[0146] Where, α M α C α L These represent the country-specific thresholds for patent market share, quality score percentage, and high-quality patent market share within the key core group, respectively; ∧ indicates that all three conditions are met simultaneously.

[0147] It should be noted that after obtaining the risk technology groups to be explored in step S4, leading institution verification is performed. The entropy method is used to assign weights to the strength indicators of each selected patent application agency, calculating the strength score of each patent application agency within each risk technology group. The patent application agency with the highest strength score within each risk technology group is selected as the leading institution, used to verify the effectiveness of identifying risk technologies using patent quality indicators in step S4. The strength indicators of the patent application agencies include the total number of patents, the sum of patent quality scores, and the total number of high-quality patents for each patent application agency within each risk technology group.

[0148] The strength score for each patent application agency is calculated using the same method as the quality score for key core patent data: first, the standardized value of the indicator data is calculated; then, the entropy value of the indicator data is calculated; next, the weight of the indicator data is calculated; finally, the weights and standardized values ​​of the indicator data are weighted and summed to calculate the strength score. Therefore, the leading agency Apps within the vth risk technology group... v The methods for obtaining it are as follows:

[0149]

[0150] Where N_APPS is the set threshold for the number of leading institutions; argmax N_APPS The top N_APPS patent application agencies have the largest values; A(app) represents the strength score of the app-th patent application agency; Apps represents the number of patent application agencies in the v-th risk technology group.

[0151] 5. Patent technology phrase extraction and text clustering ranking

[0152] S5. After pre-designing the schema constraints, use the fine-tuned information extraction model to extract fine-grained technical phrases from the patent texts within the risk technology group. Then, cluster the extracted fine-grained technical phrases into a set of technical phrase clusters, taking the risk technology group as the unit. Calculate and rank the quality scores of each technical phrase cluster in the set. Select the technical phrase clusters with the highest quality scores in each risk technology group as representative technical phrases for experts to analyze and verify the risk technologies in the patent data.

[0153] It should be noted that step S5 of this invention consists of three sub-steps: technical phrase extraction and calculation, phrase text clustering, and phrase ranking and filtering. The specific process is as follows:

[0154] S51. Technical phrase extraction and calculation: Pre-acquire a batch of labeled patent data. Where, x txt For the original text, ytxt The list of labeled technical phrases is provided, and TXT represents the total number of labeled texts. The pre-trained information extraction model M(θ) is fine-tuned for entity extraction using labeled patent data. This entity extraction fine-tuning involves minimizing the loss function L(θ; D) to fine-tune the parameters θ of the information extraction model.

[0155]

[0156] θ * This represents the parameters of the information extraction model after fine-tuning entity extraction.

[0157] S52. Extract patent texts from the risk technology group to be explored. And the schema constraints are input into the information extraction model M(θ) after entity extraction fine-tuning. * Extract fine-grained technical phrases from the patent text.

[0158] S53. Fine-grained technical phrases extracted from the same patent text are processed according to the information extraction model M(θ). * The given probability sorting process is as follows:

[0159]

[0160] in, These represent the 1st, 2nd, ..., nsth fine-grained technical phrases, respectively. Let represent the probabilities of the 1st, 2nd, ..., nsth fine-grained technical phrases given the patent text and information extraction model, respectively.

[0161] S54. Select the top 10 fine-grained technical phrases with the highest probability from the sorted fine-grained technical phrases as the patent text for the risk technology group. List of technical phrases Here, Phrase represents the preset threshold for the number of fine-grained technical phrases.

[0162] S55. Phrase Text Clustering: After obtaining the corresponding list of technical phrases for each patent text in each risk technology group, hierarchical clustering is performed on the fine-grained technical phrases in all technical phrase lists within each risk technology group, and the fine-grained technical phrases in all technical phrase lists within each risk technology group are converted into text vectors. The cosine similarity between any two text vectors is calculated, and the distance metric between the two text vectors is obtained after converting the cosine similarity calculation result. Clustering stops when the minimum distance between the text vectors of any two clusters exceeds the preset text vector threshold, and a set of technical phrase clusters is obtained.

[0163] The set of technical phrase clusters C within the rg-th risk technology group rg Represented as:

[0164]

[0165] Among them, C r C g ∈C rg Representing the set of technical phrase clusters C rg The text contains two distinct technical phrase clusters; a1, b1, a2, b2 represent arbitrary fine-grained technical phrases within the corresponding technical phrase clusters; d(a1, b1) and d(a2, b2) both represent the distance metric between two text vectors; β represents a preset text vector threshold; ∧ represents simultaneous satisfaction; r and g represent the set of technical phrase clusters C. rg Indexes of two different technical phrase clusters.

[0166] It should be noted that in step S5 of this embodiment, the bag-of-words model is used to calculate the cosine similarity between text vectors. All technical phrases within each risk technology group are segmented using an open-source toolkit to construct a list of technical phrases. For fine-grained technical phrases in the list of technical phrases within each risk technology group, each fine-grained technical phrase is converted into a text vector v using any appropriate model. count Text vector v count The j-th element v count, This represents the j-th word in the list of technical phrases. j In fine-grained technology phrases Number of times it appears in:

[0167] v count =(v count,1 ,v count,2 ,…,v count,n )

[0168]

[0169] Among them, v count,1 ,v count,2 ,…,v count,n These represent text vectors v respectively. count The 1st, 2nd, ..., nth elements. Hierarchical clustering algorithm is used to cluster the technical phrase vectors, with a preset text vector threshold β. The distance metric is calculated based on the cosine similarity result; a higher cosine similarity value indicates that the two text vectors are closer. Therefore, subtracting the cosine similarity result from 1 yields the distance metric between the two text vectors.

[0170] S56. Phrase sorting and filtering: Calculate the score of each technical phrase cluster within each risk technical group, and select the technical phrase cluster with the highest score as the representative technical phrase of the risk technical group.

[0171] The rg-th risk technology group, the r-th technology phrase cluster C r Score The calculation formula is as follows:

[0172]

[0173] Where NUM represents the total number of patents in the rg-th risk technology group; T t S represents the list of technical phrases in the t-th patent within the rg-th risk technology group; t This represents the quality score of the t-th patent within the rg-th risk technology group; ∩ represents the empty set; ∩ represents the intersection of the sets.

[0174] The representative technical phrase of the rgth risk technology group is RepTech. rg as follows:

[0175]

[0176] Where Kk is the set threshold for the number of phrases, and argmax Kk This represents the top Kk technical phrase clusters with the highest scores.

[0177] To better demonstrate the specific implementation and technical effects of the present invention, the risk technology identification method based on patent data shown in steps S1 to S5 of the above preferred implementation is applied to a specific example.

[0178] Example

[0179] In this embodiment, the risk technology identification method based on patent data is implemented as shown in steps S1 to S5, and will not be repeated here. Only the parameter settings and data results during the implementation of this invention are shown.

[0180] In step S1, the technical keywords for the design are shown in Table 1:

[0181] Table 1. Examples of Technical Keywords

[0182]

[0183]

[0184] In step S1 of this embodiment, the relevant classification numbers of the "Key Digital Technology Patent Classification System (2023)" document related to the integrated circuit field are shown in Table 2:

[0185] Table 2. Examples of Patent Classification Numbers

[0186]

[0187]

[0188] Based on the technical keywords and patent classification numbers in S1, a total of 190,064 patents in the field of integrated circuits were retrieved. Among the patent data to be mined obtained in step S1, 75,822 patents had a citation count greater than 0, which was used to construct the IPC technology co-occurrence network in step S2. The patent classification numbers in the patent data to be mined contained 9,935 IPC technology groups, generating 190,461 edges. The maximum edge weight was 3,357, and the minimum weight was 1. In step S3, when calculating node centrality, the node quantity percentage threshold b was set to 0.01, and the node degree threshold a was set to 1,000. Finally, the top c = 99 nodes with the highest node centrality values ​​were selected from the IPC technology co-occurrence network point set. Table 4 shows some IPC technology groups and their centralities in the first original node set DegSet.

[0189] Table 4. Examples of some IPC technical groups and node centrality

[0190]

[0191]

[0192] In step S3, K-core analysis is performed on the IPC technology co-occurrence network to obtain k. max =232, meaning that the degree of each node in the subgraph is greater than or equal to 232. The number of nodes in the subgraph is 179. The results of the node centrality calculation are shown in Table 5.

[0193] Table 5. Results of node centrality calculation

[0194] 1 H01L29 / 66 18172 2 H01L29 / 78 16919 3 G11C16 / 04 13471 ...... ...... ...... 178 H01L27 / 11517 243 179 H01L21 / 22 232

[0195] In step S3, when recalling patents, the intersection of the first original node set and the second original node set is taken to obtain a total of 82 IPC technical groups as candidate key core groups. A total of 83,168 patents with these IPC technical groups as the main classification number are recalled.

[0196] In step S4 of this embodiment, when calculating the quality score of each key core patent data item, there are a total of 83,168 key core patent data items and 10 indicators for measuring patent quality. In this embodiment, the weights of the 10 patent quality indicators are shown in Table 6:

[0197] Table 6. Examples of Indicator Weights for Patent Quality

[0198]

[0199]

[0200] Based on the weights in Table 6, the quality scores of the key core patent data in the recall were calculated, and the quality scores of the key core patents are shown in Table 7:

[0201] Table 7. Quality Score Calculation Results

[0202] 1 US5731856A 4074 0.369639899 2 US7211825B2 3881 0.350000448 3 US7402506B2 3865 0.349942465 4 US7453087B2 3675 0.323951427 ...... ...... ...... ...... 83167 TW508513B 0 0 83168 JP06012802B2 0 0

[0203] In step S4 of this embodiment, the target country is set to China, and the country-specific gap threshold α of the patent market share within the key core group is taken. M = -0.3, the threshold α for country-specific differences in quality score percentage C = -0.3, Country-specific threshold α for market share of high-quality patents L = -0.3. Through country-specific gap analysis, in 9 out of the 82 key core groups, China's gap with the top-ranked country in terms of patent quantity, patent quality, and number of high-quality patents exceeded the set threshold. The results are shown in Table 8.

[0204] Table 8. Results of Country-Specific Gap Analysis

[0205] H01L21 / 822 -0.743 -0.983 -0.702 H01L27 / 10 -0.489 -0.988 -0.90 H01L29 / 94 -0.416 -0.991 -0.769 H01L27 / 118 -0.73 -0.99 -0.333 H01L21 / 338 -0.631 -0.962 -0.714 H01L21 / 316 -0.395 -0.964 -0.889 H01L21 / 3205 -0.417 -0.985 -0.607 G02F1 / 136 -0.681 -0.946 -0.67 H01L27 / 04 -0.592 -0.972 -0.506

[0206] In this embodiment, after completing step S4, the leading institution verification begins. The strength scores of the leading institutions in the nine risk technology groups from the previous stage are calculated, with the threshold for the number of leading institutions, N_APPS, set to 3. The results are shown in Table 9.

[0207] Table 9. Leading Institutions and Strength Scores

[0208]

[0209] In step S5 of this embodiment, when fine-tuning the parameters of the information extraction model, a total of 200 patent data entries were labeled, and the designed schema constraints were as follows:

[0210] [“technology”, “device”, “equipment”, “device”, “system”, “method”, “software”, “material”]

[0211] The information extraction model adopts the unified information extraction framework model UIE-base, based on the ERNIE 3.0 pre-trained model, and is fine-tuned for entity extraction tasks. The loss function used is Loss. FTas follows:

[0212]

[0213] Where, D={(schema,x txt ,y txt )} represents the labeled patent data, θ e θ d These are the parameters for the encoder (Econder) and decoder (Decoder), respectively.

[0214] For fine-tuning, set batch_size=10 and epoch=50. After fine-tuning, use the above schema constraints to extract fine-grained technical phrases from the patent texts contained in the risk technology group, and sort the technical phrases of the patent texts according to the extraction probability given by UIE-base. The threshold is set to 3, that is, only 3 technical phrases with the highest probability are retained for each patent.

[0215] In step S5 of this embodiment, during phrase text clustering, the text vector threshold β is set to 0.4. Phrase-level clustering is performed on the nine risk technology groups obtained in the previous step, and the results are shown in Table 10.

[0216] Table 10. Clustering Results

[0217] H01L21 / 822 76 phrases 12 clusters H01L27 / 10 71 phrases 14 clusters H01L29 / 94 26 phrases and 5 clusters H01L27 / 118 33 phrases 5 clusters H01L21 / 338 79 phrases, 11 clusters H01L21 / 316 15 phrases 3 clusters H01L21 / 3205 43 phrases 6 clusters G02F1 / 136 96 phrases 8 clusters H01L27 / 04 132 phrases 21 clusters

[0218] In step S5 of this embodiment, during the phrase sorting and filtering process, the phrase quantity threshold Kk is set to 3. Within each risk technology group in the previous step, the technical phrase cluster scores are statistically analyzed and sorted. The representative technical phrase descriptions and scores of the risk technology groups are shown in Table 11.

[0219] Table 11. Descriptions and Scores of Representative Technical Phrases

[0220]

[0221]

[0222] Analysis of the results reveals that China lags significantly behind other countries in the fields of semiconductor devices and materials (non-volatile memory, group III nitride semiconductor transistors, gallium nitride semiconductor devices, thin-film transistors), circuit design (electrostatic discharge protection components, clock drive circuits, MOS transistor components), manufacturing and packaging (hybrid layered superlattice materials, thin-film transistors, active matrix liquid crystal display devices, chemical mechanical polishing methods), and memory technology (non-volatile memory, semiconductor memory, asymmetric semiconductor memory cells) within the integrated circuit industry. These areas are indeed crucial components of integrated circuits and represent the most intense technological competition currently. The results demonstrate that the method of this invention helps identify potential technological fields for a nation, offering forward-looking insights for independent technological innovation and breakthrough technological challenges.

[0223] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the invention. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the invention. Therefore, all technical solutions obtained through equivalent substitution or transformation fall within the protection scope of the present invention.

Claims

1. A method for identifying risky technologies based on patent data, characterized in that, Includes the following steps: S1. Obtain technical keywords within the target industry sector. After translating and performing synonym substitution on the technical keywords, search the global patent database for patents whose titles or abstracts contain the technical keywords. Obtain the patent classification number within the target industry sector according to the technology classification system of the target industry sector. Obtain the corresponding patents according to the patent classification number. Combine the patents retrieved through technical keywords and the patents obtained according to the patent classification number as the patent data to be mined in the target industry sector. The target industry sector is the industry sector that needs to be mined for technology risk. S2. Construct an IPC technology co-occurrence network using the patent data to be mined obtained in step S1. The IPC technology co-occurrence network is a graph network consisting of nodes, edges, and edge weights. The nodes are IPC technology groups that have appeared in the patent classification numbers of the patent data to be mined. The edges represent the co-occurrence relationship between two IPC technology groups. The weight of the edge is the number of times the two IPC technology groups co-occur in the patent classification numbers of the patent data to be mined. S3. Perform node centrality statistics on each node in the IPC technology co-occurrence network, and construct the first original node set using the nodes with the highest node centrality ranking in the IPC technology co-occurrence network; perform K-core analysis on the IPC technology co-occurrence network to obtain the maximum K-core subgraph, and construct the second original node set using the nodes in the maximum K-core subgraph; take the intersection of the first original node set and the second original node set to obtain the key core node set, and take the IPC technology groups in the key core node set as the key core groups; recall all patents in the patent data to be mined that use the key core groups as the main IPC classification numbers, and take them as the key core patent data to be mined; S4. After pre-constructing indicators to measure patent quality based on the characteristics of risky technologies, calculate the quality score of all key core patent data. In each key core group, the core patent data with the highest quality score are regarded as high-quality patents for each key core group. The key core groups that meet the preset requirements of the evaluation indicators are regarded as risk technology groups to be explored. The evaluation indicators are the patent market share, quality score ratio and national gap of high-quality patent market share within the key core group. S5. After pre-designing the schema constraints, use the fine-tuned information extraction model to extract fine-grained technical phrases from the patent texts within the risk technology group. Then, cluster the extracted fine-grained technical phrases into a set of technical phrase clusters, taking the risk technology group as the unit. Calculate and rank the quality scores of each technical phrase cluster in the set. Select the technical phrase clusters with the highest quality scores in each risk technology group as representative technical phrases for experts to analyze and verify the risk technologies in the patent data. The specific process of step S3 is as follows: S31. The centrality of a node is the sum of the weights of its neighboring edges, calculated using the following formula: ； in, For IPC technology co-occurrence network point sets Middle node The sum of the weights of adjacent edges; Represents the network point set co-occurring with IPC technology In and nodes A connected node; For the edge set of the IPC technology co-occurring network; Represents a node and nodes The weights of the edges formed; Represents the edge set of a co-occurring network of IPC technology From the node and nodes A single edge is formed; S32. In the IPC technology co-occurrence network, select the nodes with the highest node centrality ranking to construct the first original node set. The calculation process is as follows: ； ； in, Represents the network point set co-occurring with IPC technology The number of nodes in; The set percentage threshold for the number of nodes; The set threshold for node degree; This indicates the network point set co-occurring in IPC technology. The node with the largest centrality value is the first One node; S33, in the K-core analysis of the first In this analysis, the sum of the weights of all neighboring edges in the IPC co-occurrence network is less than [a certain value]. The nodes and their connected edges are identified to obtain the original subgraph. The weights of the adjacent edges of the remaining nodes in the original subgraph are then recalculated until no adjacent edge in the original subgraph has a weight less than 1. The node, to obtain the first The subgraph of the second analysis; where, the first The set of nodes in the subgraph analyzed in this step for: ； In the formula, Indicates the first Nodes in the subgraph analyzed below The sum of the weights of the adjacent edges; This represents the subgraph of the k-th analysis. S34. Continue the analysis until the... The analysis revealed the first IPC technology co-occurrence network. When the subgraph in the next analysis is empty, As the maximum number of times K-core analysis is performed, the th The subgraph analyzed in this step is the largest K-kernel subgraph of the IPC technology co-occurrence network; S35. Take the set of nodes in the maximum K-kernel subgraph as the second original node set, and the first original node set... With the second original node set Taking the intersection yields the set of key core nodes. ; S36. Recall all patents in the patent data to be mined that use the IPC technical group in the set of key core nodes as the IPC main classification number, and use them as key core patent data to be mined.

2. The risk technology identification method based on patent data as described in claim 1, characterized in that, In step S1, the technology classification system of the target industry field refers to the official document that includes the technical field and the reference relationship with the international patent classification system.

3. The risk technology identification method based on patent data as described in claim 1, characterized in that, The specific process of step S4 is as follows: S41. Using the entropy method, weights are assigned to the selected patent quality indicators, and the quality scores of all key core patent data are calculated; if there are a total of key core patent data... There are a total of [number] indicators for patent quality. The quality score of the i-th key core patent data item. The calculation process is as follows: ； in, The weight of the indicator representing the quality of the j-th patent; The standardized value of the indicator representing the quality of the i-th key core patent data and the j-th patent. S42. Taking key core groups as units, a pre-set percentage threshold p is set, and the patents with the highest quality scores in each key core group are regarded as high-quality patents of the key core group, and the evaluation indicators of each key core group are calculated. The formula for calculating the country-specific differences in patent market share within key core groups is as follows: ； in, The classification number is The gap in patent market share between the target countries within the key core group and the country with the highest number of patent applications; This indicates that the target country is in the classification number. The number of patents within the key core group; The country with the highest number of patent applications is classified as follows: The number of patents within the key core group; For the classification number The total number of countries applying for patents within the key core group; Representing the Patent application countries in classification number The number of patents within the key core group; The formula for calculating the country-specific differences in the percentage of quality scores within key core groups is as follows: ； in, The classification number is The gap in patent quality score percentage between the target country within the key core group and the top-ranked patent application country; , , These respectively represent the target country, the country with the highest number of patent applications, and all countries with patent applications in the classification number. The total number of patents within the key core group; , , These represent the target country, the country with the highest number of patent applications, and the country ranked [number] among all patent application countries. Quality scores of key core patent data; The formula for calculating the country-specific differences in market share of high-quality patents within key core groups is as follows: ； in, The classification number is The gap between the target countries within the key core group and the top-ranked patent applicant country in terms of market share of high-quality patents; This indicates that the target country is in the classification number. The number of high-quality patents within the key core group; This indicates that the country with the highest number of patent applications is in the classification number. The number of high-quality patents within the key core group; Indicates the first Patent application countries in classification number The number of high-quality patents within the key core group; S44. Determine whether the three evaluation indicators of each key core group simultaneously meet the set thresholds corresponding to each evaluation indicator: if they simultaneously meet the thresholds, then the key core group is identified as a risk technology group to be explored.

4. The risk technology identification method based on patent data as described in claim 3, characterized in that, In step S4, the weight of the j-th patent quality index The calculation method is as follows: ； in, The entropy value, representing the quality index of the j-th patent, is calculated as follows: ； in, This indicates finding the reciprocal.

5. The risk technology identification method based on patent data as described in claim 3, characterized in that, In step S4, the standardized value of the quality index of the i-th key core patent data item j-th patent is... The calculation method is as follows: ； in, These represent the minimum and maximum values ​​of the quality index of the j-th patent among all key core patent data.

6. The risk technology identification method based on patent data as described in claim 1, characterized in that, In step S4, a total of 10 indicators for measuring patent quality are selected, namely, the number of patent families, the number of countries where patent families are deployed, the number of countries where patents are cited, the average annual citation frequency, the number of citations, the coverage of IPCs, the number of inventors, the number of countries of invention, the number of citations, and the number of claims.

7. The risk technology identification method based on patent data as described in claim 1, characterized in that, After obtaining the risk technology groups to be explored in step S4, the entropy method is used to assign weights to the strength indicators of each selected patent application agency, and the strength score of each patent application agency in each risk technology group is calculated. The patent application agency with the highest strength score in each risk technology group is selected as the leading agency to verify the effectiveness of identifying risk technologies using patent quality indicators in step S4. The strength indicators of the patent application agency are the total number of patents, the sum of patent quality scores, and the total number of high-quality patents of the patent application agencies in each risk technology group. No. Leading institutions within the risk technology group The methods for obtaining it are as follows: ； in, The threshold for the number of leading institutions is set; Represents the largest value among the top values. One patent application agency; Indicates the first The strength score of each patent application agency; Indicates the first The number of patent application agencies within each risk technology group.

8. The risk technology identification method based on patent data as described in claim 1, characterized in that, The specific process of step S5 is as follows: S51. Pre-acquire a batch of labeled patent data. Using labeled patent data to train a pre-trained information extraction model Perform entity extraction fine-tuning; among which, This is the original text. A list of technical phrases with annotations. The total number of texts that are annotated; S52. Extract patent texts from the risk technology group to be explored. And schema constraints are input into the information extraction model after entity extraction fine-tuning. In the process, fine-grained technical phrases are extracted from the patent text. ; S53. Fine-grained technical phrases extracted from the same patent text are processed according to an information extraction model. The given probability sorting, specifically the sorting process is as follows: ； ； in, They represent the first A fine-grained technical phrase; They represent the first The probability of a fine-grained technical phrase given a patent text and an information extraction model; S54. Select the top 10 fine-grained technical phrases with the highest probability from the sorted fine-grained technical phrases as the patent text for the risk technology group. List of technical phrases Where Phrase represents the preset threshold for the number of fine-grained technical phrases; S55. After obtaining the corresponding list of technical phrases for each patent text in each risk technology group, hierarchical clustering is performed on the fine-grained technical phrases in all technical phrase lists within each risk technology group, and the fine-grained technical phrases in all technical phrase lists within each risk technology group are converted into text vectors. The cosine similarity between any two text vectors is calculated, and the distance metric between the two text vectors is obtained after converting the cosine similarity calculation result. Clustering is stopped when the minimum distance between the text vectors of any two clusters exceeds the preset text vector threshold, and a set of technical phrase clusters is obtained. The set of technical phrases within the rgth risk technology group Represented as: ； in, Each represents a set of technical phrase clusters. Two different clusters of technical phrases; Each represents an arbitrary fine-grained technical phrase within the corresponding technical phrase cluster; Both represent the distance metric between two text vectors; This represents the preset text vector threshold. This means that both conditions are met simultaneously; Represents a set of technical phrase clusters Indexes of two different technical phrase clusters; S56. Calculate the score of each technical phrase cluster within each risk technical group, and select the technical phrase cluster with the highest score as the representative technical phrase of the risk technical group. No. The first risk technology group A cluster of technical phrases Score The calculation formula is as follows: ； in, Indicates the first The total number of patents within each risk technology group; Indicates the first List of technical phrases for patent t within each risk technology group; Indicates the first Quality score of the t-th patent within the risk technology group; Represents the empty set; This indicates finding the intersection; No. Representative technical phrases of each risk technology group as follows: ； in, For the set threshold number of phrases, The top-ranked technical phrase cluster score represents the highest value among all clusters of technical phrases. A cluster of technical phrases.

Images