Patent multi-dimensional evaluation method and system based on large language model

Abstract

The invention provides a patent multi-dimensional evaluation method and system based on a large language model, and belongs to the technical field of literature analysis. According to the scheme, a patent cluster to which each patent belongs is determined through a technical cluster constructed based on clustering analysis and the large language model, and the patent multi-dimensional evaluation method and system based on the large language model are established through a mode of constructing a patent cluster network. According to the method, the importance of each patent in the whole patent cluster is determined through the multi-dimensional features of the patent cluster, then the patent is evaluated in combination with the multi-dimensional features of the patent cluster, original patent data is effectively utilized, and the comprehensiveness and accuracy of the patent evaluation result are ensured; the method effectively solves the problems that the accuracy of a clustering analysis result of a traditional scheme cannot be guaranteed, the correlation between patents in the same field and cross-field is not considered, the accuracy of a patent evaluation result is greatly influenced, the evaluation dimension of an existing patent evaluation mode is single, and the evaluation efficiency is low. And the accuracy of the patent evaluation result is insufficient.

Description

Technical Field

[0001] The present invention belongs to the technical field of document analysis, and in particular, relates to a patent multi-dimensional evaluation method and system based on a large language model. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] Most traditional patent evaluation schemes adopt a subjective judgment method and ignore the use of original patent data. Secondly, although some patent evaluation schemes use patent clustering to analyze and evaluate patents, they usually use some conventional clustering algorithms, resulting in poor accuracy of clustering results. It is easy to classify patents with large differences in fields into the same category or have more outliers, resulting in distortion of clustering results and inability to obtain accurate clustering analysis results. If patent evaluation is performed based on the results of the above cluster analysis, the accuracy of the clustering analysis results cannot be guaranteed, and there is a lack of consideration of the correlation between patents in the same field and across fields, which will have a greater impact on the accuracy of the patent evaluation results. At the same time, the existing patent evaluation methods have a single evaluation dimension, which makes the accuracy of the patent evaluation results insufficient. Summary of the invention

[0004] The present invention provides a patent multi-dimensional evaluation method and system based on a large language model to solve the problem that in traditional solutions, in patent clustering analysis, patents with large field differences are easily classified into the same category or a large number of outliers appear, resulting in distortion of clustering results and inability to obtain accurate analysis results; and in patent evaluation, since the accuracy of clustering analysis results cannot be guaranteed and there is a lack of consideration of the correlation between patents in the same field and across fields, which will have a significant impact on the accuracy of patent evaluation results; and the existing patent evaluation method has a single evaluation dimension, resulting in insufficient accuracy of patent evaluation results.

[0005] According to a first aspect of an embodiment of the present invention, a patent multi-dimensional evaluation method based on a large language model is provided, comprising: Extract keywords from the patents in the pre-built patent raw data set, perform clustering based on the extracted keywords, and obtain several initial technology clusters; For each initial technology cluster obtained, optimize it using a pre-trained large language model to obtain an optimized technology cluster; Based on the keywords extracted from the patents and combined with the optimized technology clusters, the technology cluster to which the current patent belongs is determined, and the patents belonging to the same technology cluster are regarded as a patent cluster to realize the construction of patent clusters; The patent cluster network is constructed by taking the patents in the patent cluster as nodes and the similarities between patents as edges. Based on the constructed patent cluster network, the importance score of each patent is calculated. Based on the importance score of the patent, combined with the number of applicants, the number of patents, the number of inventors of the cluster to which the patent belongs, and the geographical distribution of the patent, the evaluation results of the patent are obtained.

[0006] Furthermore, the patents in the pre-constructed original patent data set are subjected to keyword extraction, and clustering processing is performed based on the extracted keywords to obtain a number of initial technology clusters, specifically: the patents in the original patent data set are divided into a number of categories based on the patent IPC classification number; keyword extraction is performed on the patents in each category; based on the extracted keywords, clustering results in each category are obtained through cluster analysis; based on the obtained clustering results in each category, a number of initial technology clusters are obtained through secondary clustering.

[0007] Furthermore, based on the extracted keywords, clustering analysis is performed to obtain clustering results under each category, specifically: clustering the keywords under the current category based on a preset clustering algorithm to obtain a first clustering set and several unclustered isolated keywords; clustering the isolated keywords and the keywords in the largest cluster in the first clustering set using a preset clustering algorithm to obtain a second clustering set; merging the first clustering set and the second clustering set to obtain a clustering result.

[0008] Furthermore, based on the clustering results obtained under each category, several initial technology clusters are obtained through secondary clustering. Specifically, based on the clustering results obtained under each category, the average vector representation of each member in each cluster is used as the feature representation of the cluster. Based on the feature representation of each cluster, secondary clustering is performed using a preset clustering algorithm. Based on the obtained clustering results, several technology clusters corresponding to the entire patent original data set are obtained.

[0009] Furthermore, based on the constructed patent cluster network, the importance score of each patent is calculated. Specifically, the PageRank algorithm is used to calculate the PageRank value of each node in the patent cluster network, and the PageRank value is used as the importance score of the patent.

[0010] Furthermore, the importance score is calculated using the following formula: ; in, is the PageRank value of patent v, d is the damping coefficient, N is the total number of patents in the cluster, u, v represents a patent in the cluster, is the set of all patents whose cosine similarity with patent v is higher than the set threshold, is the number of cosine similarities between patent v and other patents that are above the threshold.

[0011] Furthermore, the similarity between the patents is calculated based on the similarity of two patent feature vectors, wherein the patent feature vector is obtained by weighted summation of several keyword vectors extracted from the patents, and the weight of each keyword is determined based on the TF-IDF algorithm.

[0012] Furthermore, the technology cluster to which the current patent belongs is determined as follows: based on the keywords corresponding to the patent, the technology clusters to which these keywords belong are queried. If the technology cluster is found, the current patent belongs to the technology cluster; if the technology cluster is not found, the similarity between the patent keywords and each technology cluster is calculated, and the technology cluster to which the current patent belongs is determined based on the similarity results.

[0013] Furthermore, the patent importance score is combined with the number of applicants, the number of patents, the number of inventors and the geographical distribution of the patents in the cluster to which the patent belongs to obtain the evaluation results of the patent, which are specifically expressed as follows: ; ; ; ; in, is the logarithmic function with base e, , , , and They are the weights corresponding to the importance score of patents, the number of patents, the number of applicants, the number of inventors, and the geographical distribution of patents; , , , and They respectively represent the patent importance score, the number of patents, the number of applicants, the number of inventors, and the geographical distribution of patents.

[0014] According to a second aspect of an embodiment of the present invention, a patent multi-dimensional evaluation system based on a large language model is provided, comprising: An initial technology cluster construction unit, which is used to extract keywords from patents in a pre-constructed original patent data set, perform clustering processing based on the extracted keywords, and obtain a number of initial technology clusters; A technology cluster optimization unit, which is used to optimize each obtained initial technology cluster using a pre-trained large language model to obtain an optimized technology cluster; Patent cluster construction unit, which is used to determine the technology cluster to which the current patent belongs based on the keywords extracted from the patent and the optimized technology cluster, and to regard the patents belonging to the same technology cluster as a patent cluster to realize the construction of patent cluster; The importance scoring unit is used to construct a patent cluster network using the patents in the patent cluster as nodes and the similarities between patents as edges; based on the constructed patent cluster network, the importance score of each patent is calculated; The patent evaluation unit is used to obtain the evaluation results of the patent based on the importance score of the patent, combined with the number of applicants, the number of patents, the number of inventors and the geographical distribution of the patent in the cluster to which the patent belongs.

[0015] Compared with the prior art, the present invention has the following beneficial effects: The solution described in the present invention proposes a patent multi-dimensional evaluation method and system based on a large language model. The solution determines the patent cluster to which each patent belongs by constructing a technology cluster based on cluster analysis and a large language model, and determines the importance of each patent in the entire patent cluster by constructing a patent cluster network. At the same time, combined with the number of applicants, the number of patents, the number of inventors and the geographical distribution of patents in the cluster to which the patent belongs, the patent is evaluated from multiple dimensions, the original patent data is effectively utilized, and the comprehensiveness and accuracy of the patent evaluation results are ensured.

[0016] The solution described in the present invention uses a large language model to optimize the technology cluster, which not only provides an efficient and universal processing solution for data governance and optimization of the technology cluster, but also effectively improves the efficiency and quality of the optimization processing.

[0017] The solution described in the present invention effectively combines the TF-IDF algorithm with the PageRank algorithm. The TF-IDF algorithm is used to strengthen the importance of the patent's keywords in the document, and the PageRank algorithm is used to evaluate the influence and core degree of the patent in its patent cluster. Combined with the multi-dimensional characteristics of the number of applicants, the number of patents, the number of inventors, and the geographical distribution of patents in the cluster to which the patent belongs, the accuracy and effectiveness of patent evaluation can be further improved, and the problem of insufficient accuracy of patent evaluation caused by traditional methods relying on subjective judgment can be effectively solved.

[0018] The solution of the present invention can effectively reflect the relative importance of each patent in the entire patent cluster network by constructing a patent cluster network and scoring patent importance through the PageRank algorithm. It is a comprehensive result based on the statistical representation of keywords and the mutual connections between patent documents. It not only starts from the content analysis of a single document, but also combines the mutual relationships between patents to comprehensively evaluate the importance and influence of patents.

[0019] When constructing technology clusters, the solution described in the present invention explores and integrates technology clusters under different IPC classifications through multiple clustering, which can comprehensively capture the technical connections between patents in the same field and across fields, and effectively improve the overall quality and accuracy of clustering.

[0020] Advantages of additional aspects of the present invention will be given in part in the following description, and in part will become obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0022] Figure 1 It is a flow chart of a patent multi-dimensional evaluation method based on a large language model described in an embodiment of the present invention; Figure 2 is a flowchart of keyword extraction described in an embodiment of the present invention; Figure 3 Construct a flow chart for the technology cluster described in the embodiment of the present invention; Figure 4 It is a flowchart of the technology cluster optimization described in the embodiment of the present invention; Figure 5 A flowchart for building a patent cluster described in an embodiment of the present invention. DETAILED DESCRIPTION

[0023] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0024] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0025] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0026] In the absence of conflict, the embodiments of the present invention and the features of the embodiments may be combined with each other.

[0027] In one or more embodiments, Figure 1 As shown, this embodiment provides a patent multi-dimensional evaluation method based on a large language model. It should be noted that in the following embodiments, "patent" mainly refers to patents in a public state. The method specifically includes the following processing procedures: Step 1: Extract keywords from the patents in the pre-built patent raw data set, perform clustering based on the extracted keywords, and obtain several initial technology clusters; Specifically, Figure 2 As shown in the figure, the specific process of patent keyword extraction is shown, which includes the following steps: (1) The patents in the original data set were divided into eight categories according to their IPC classification codes: necessities of human life (A), operations; transportation (B), chemistry; metallurgy (C), textiles; papermaking (D), fixed buildings (E), mechanical engineering; lighting; heating; weapons; blasting (F), physics (G), and electricity (H); (2) Use a pre-trained keyword extraction model to extract keywords from the patents in each category, and use a general vector model (such as BGE-M3) to vectorize the keywords of the patents; like Figure 3 As shown in the figure, the construction process of the initial technology cluster is shown, which includes the following steps: (1) Clean the generated patent keywords and remove stop words; (2) Keywords are summarized according to the above eight categories, and clustered using the HDBSCAN method based on the vectors corresponding to the keywords in each category; The HDBSCAN algorithm is used to form clusters of technical keywords in a given vector space. The HDBSCAN algorithm is simplified as follows: ; in, is a collection of keyword vectors, is the minimum number required to form an effective cluster, which is set to 20 in the solution described in this embodiment. Indicates the distance or similarity metric used. In the solution described in this embodiment, cosine similarity is used.

[0028] It is understood that in one or more embodiments, The value of can be set according to actual needs, and the metric can also use other similarity calculation methods.

[0029] Finally, the clustering set A includes { } and several isolated keywords ; (3) Using the HDBSCAN method to classify isolated keywords that are not clustered and the keywords in the largest cluster (assuming ) re-clustering; after re-clustering, a new clustering set B is obtained including { } and isolated keywords ,calculate With the collection { } and isolate keywords Merge into the cluster with the highest similarity.

[0030] (4) In the merging operation, this embodiment uses a direct merging method to process the cluster set A and cluster set B generated above.

[0031] Specifically, merge the other sets in clustering set A except the largest cluster , and the set in cluster set B ,Since the two clusterings are independent, the labels may be repeated. When merging, the cluster labels from different sources need to be renamed to ensure the uniqueness of the labels. For example: cluster set A { }Renumbered to , the cluster set B{ }Renumbered to , and get the final initial clustering result { }.

[0032] (5) Adjust the clusters based on cosine similarity.

[0033] Specifically, clusters containing more than a preset number of keywords (e.g., 1,000) are considered "too large" and need to be split reasonably. On the contrary, clusters containing fewer than a preset number of keywords (e.g., 50) are considered "too small" and should be merged with other small clusters with higher relevance to enhance the representativeness and data density of the cluster; in this way, cluster data can be managed and utilized more effectively to ensure the quality and effectiveness of data analysis.

[0034] (6) Calculate the average vector of the clusters formed, and use HDBSCAN to perform secondary clustering; for each cluster generated by the first clustering, calculate the average vector of its members as the feature representation of the cluster, and use the HDBSCAN algorithm to perform secondary clustering on these average vectors. This process aims to discover higher-level structural relationships and further aggregate clusters with similar characteristics, thereby forming a new, larger set of clusters.

[0035] It should be noted here that the secondary clustering of the solution described in this embodiment has the following advantages: Improve clustering quality: Although the first clustering can effectively reduce the impact of noise and outliers based on IPC classification, it may not fully capture the technical associations across classifications. Through secondary clustering, we can further explore and integrate technical clusters under different IPC classifications, and improve the overall quality and accuracy of clustering.

[0036] Enhance the representativeness and generality of clusters: The first clustering often focuses on the subdivisions within a specific technology field, while the second clustering helps to identify broader themes or trends that span multiple technology fields, making the resulting clusters more representative and general.

[0037] Optimize resource allocation and decision support: The higher-level clusters formed by secondary clustering can provide more precise support for resource allocation, policy making, and technology transfer, helping decision makers better understand the overall situation of technological development.

[0038] (7) Based on the clustering results and labels, calculate the center of each cluster and classify the unclassified isolated keywords into the cluster with the highest cosine similarity to obtain the technology cluster; Step 2: For each initial technology cluster obtained, optimize it using the pre-trained large language model to obtain an optimized technology cluster; like Figure 4 As shown in the figure, the optimization process using a large language model is shown, which specifically includes the following processing steps: (1) First, determine whether the number of characters occupied by the cluster keywords is the maximum length that LLM can handle. If it exceeds, divide the cluster into multiple subsets; (2) Prompt settings, including: prompt = " ## Description: You are a data cleaning model. Given a dictionary consisting of a cluster ID and multiple keywords, please clean the keywords of the dictionary and finally output it in a string strictly in json format.

[0039] ## During the cleaning process, please strictly abide by the following principles: 1. Cleaning: Delete repeated and abnormal keywords.

[0040] 2. Technical vocabulary: Please delete keywords that are not technical vocabulary.

[0041] 3. Rename: Please rename the cluster according to the cleaned keywords.

[0042] 4. Output: Output keywords are sorted by importance, and all keywords after cleaning must be displayed.

[0043] 5. It is prohibited to output any information that is not related to the results.

[0044] ## Example Given a cluster dictionary {"cluster_id (i.e. cluster ID, e.g. 533)":["Multi-channel graph convolutional neural network", "Image synthesis", "Rapid sclerotium rapid generation", "Rapid flower cultivation and breeding", "Rapid gardening method", "Adaptive graph convolutional network model"]}; ### Output "json{"Convolutional Neural Network":["Multi-channel Graph Convolutional Neural Network","Image Synthesis","Adaptive Graph Convolutional Network Model"]}"; ## Task ### Given a cluster dictionary {"{cluster_id}":{keyword_string}} ### Output Among them, cluster_id is the cluster ID, which is used to identify different clusters during clustering; keyword_string is a string formed by keywords in the technology cluster;" (3) Put the preprocessed clusters together with the set prompt and input them into the large language model. The large model used in this embodiment is Qwen-72B. The large language model is used to clean the keywords in the above technical clusters, that is, to clean repeated and abnormal keywords, remove keywords that do not belong to technical vocabulary, and classify and rename the cleaned clusters.

[0045] Step 3: Based on the keywords extracted from the patent and combined with the optimized technology clusters, determine the technology cluster to which the current patent belongs, and treat the patents belonging to the same technology cluster as a patent cluster to achieve the construction of patent clusters; In the specific implementation, Figure 5As shown, the method of determining the technology cluster to which the current patent belongs is specifically as follows: based on the keywords corresponding to the patent, query the technology clusters to which these keywords belong. If the technology cluster is found, the current patent belongs to the technology cluster; if the technology cluster is not found, calculate the similarity between the patent keywords and each technology cluster, and determine the technology cluster to which the current patent belongs based on the similarity result.

[0046] Step 4: Use the patents in the patent cluster as nodes and the similarities between patents as edges to construct a patent cluster network; based on the constructed patent cluster network, calculate the importance score of each patent; In specific implementation, the calculation of the importance score of a patent includes the following processing steps: (1) Calculate the TF-IDF value of the keyword. TF-IDF can assign a weight to each word in the patent, which reflects the importance of the word to the patent. The term frequency (TF) part emphasizes the frequency of patent keywords in the patent, while the inverse document frequency (IDF) part emphasizes the rarity of patent keywords. The higher the TF-IDF value of a word, the more important the keyword is to the patent: TF (Term Frequency) is:

[0047] in, As a keyword in the cluster, d is a single patent, is the number of times keyword t appears in the patent is the sum of the number of times all keywords appear in patent d; DF (Document Frequency) is the number of patents containing keyword t in cluster D:

[0048] IDF(Inverse Document Frequency)

[0049] TF(t,d) is the frequency of the keyword in patent d, IDF(t,D), the inverse document frequency of keyword t (2) According to the calculated TF-IDF value, adjust the weight of each keyword in its corresponding patent vector; for a patent p, it contains n keywords , ,… , each keyword has a corresponding vector , …, , and TF-IDF weight , ,…, ,The vector of the patent can be expressed as:

[0050] (3) Use cosine similarity to calculate the similarity between patents. The specific formula is as follows:

[0051] Among them, A and B represent the vector representation of two patents respectively. represents the dot product of vectors, and is the magnitude of the vector; (4) Using the patents in the patent cluster as nodes and the similarity results between patents as edges, a patent cluster network is constructed; (5) Apply the PageRank algorithm on the constructed patent cluster network to calculate the importance score of each patent node:

[0052] in, is the PageRank value of patent v, d is the damping coefficient, N is the total number of patents in the cluster, u, v represents a patent in the cluster, is the set of all patents whose cosine similarity with patent v is higher than the set threshold, is the number of patent v's cosine similarities with other patents that are above the threshold. The PageRank of each patent is set to 1 in the initial stage, and then stabilizes after multiple iterations. This is generally calculated by iterative update formula until the PageRank values ​​of all nodes in the system no longer change significantly.

[0053] The PageRank algorithm is applied to the constructed network to calculate the importance score of each patent node. This score reflects the relative importance of each patent in the entire patent network. It is a comprehensive result based on the statistical representation of keywords and the interconnections between patent documents. This can not only start from the content analysis of a single document, but also combine the interrelationships between patents to comprehensively evaluate the importance and influence of patents. A high-scoring patent usually means that it has a high degree of similarity with multiple other patents, which may indicate that it belongs to a core technology field or trend.

[0054] Step 5: Based on the importance score of the patent, combined with the number of applicants, number of patents, number of inventors and geographical distribution of the patent cluster, the evaluation results of the patent are obtained.

[0055] For a patent to be evaluated, the solution described in this embodiment not only uses the importance score of the patent in the technology cluster, but also designs a multi-dimensional patent evaluation strategy based on the following core indicators: patent importance score, the number of applicants in the same cluster, the number of patents in the same cluster, the number of inventors in the same cluster, and the distribution of patents in the cluster. These indicators work together to reflect the importance and popularity of patents in the technical field.

[0056] The following is an explanation of the above indicators: Number of applicants: The number of individuals or institutions involved in patent applications in the cluster. The larger the number, the higher the competition and activity in the technology field.

[0057] Number of patents: The total number of patents in the cluster, reflecting the R&D activity in this technology field.

[0058] Number of inventors: The number of independent inventors involved in different patents in the cluster. Diversity and a wide range of participants can increase the innovation of technology. If an inventor has multiple patents in a cluster, we only count it once to measure the innovation potential of a patent cluster. We believe that the technology in a cluster with more independent inventors is more important and popular.

[0059] Exemplary: Assume we have a patent cluster containing the following patents: Patent A, inventors: Zhang San, Li Si; Patent B, inventors: Li Si, Wang Wu; Patent C, inventors: Zhao Liu, Zhang San. In this cluster, there are a total of the following inventors: Zhang San, Li Si, Wang Wu, Zhao Liu; although some inventors appear in multiple patents, we only care about the number of independent individuals. Therefore, the number of independent inventors is 4.

[0060] Patent distribution: This indicator considers the distribution width of patents in different regions. Wider geographical coverage indicates the popularity and market acceptance of the technology. From the global distribution of the patent cluster, we can analyze the development direction and business strategy of the technology. Collect and organize the patent registration addresses in the cluster, and calculate how many regions the patents in the cluster cover, as shown in the formula As shown, the more areas a cluster covers, the more it indicates that this direction is a popular direction to some extent.

[0061] Specifically, the multi-dimensional patent evaluation strategy is specifically expressed as follows:

[0062] in:

[0063]

[0064]

[0065] in, The logarithmic function with base e is used to avoid the mathematical error caused by log(x) when x=0. The use of logarithmic transformation can balance the influence of different variables so that variables with large magnitude differences can also reasonably contribute to the total score. Set to 0.6, , These indicators are set to 0.1.

[0066] In one or more embodiments, corresponding to the above method, this embodiment provides a patent multi-dimensional evaluation system based on a large language model, including: An initial technology cluster construction unit, which is used to extract keywords from patents in a pre-constructed original patent data set, perform clustering processing based on the extracted keywords, and obtain a number of initial technology clusters; A technology cluster optimization unit, which is used to optimize each obtained initial technology cluster using a pre-trained large language model to obtain an optimized technology cluster; Patent cluster construction unit, which is used to determine the technology cluster to which the current patent belongs based on the keywords extracted from the patent and the optimized technology cluster, and to regard the patents belonging to the same technology cluster as a patent cluster to realize the construction of patent cluster; The importance scoring unit is used to construct a patent cluster network using the patents in the patent cluster as nodes and the similarities between patents as edges; based on the constructed patent cluster network, the importance score of each patent is calculated; The patent evaluation unit is used to obtain the evaluation results of the patent based on the importance score of the patent, combined with the number of applicants, the number of patents, the number of inventors and the geographical distribution of the patent in the cluster to which the patent belongs.

[0067] Furthermore, the system described in this embodiment corresponds to the method described in Embodiment 1, and its technical details have been described in detail in Embodiment 1, so they will not be repeated here.

[0068] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A multi-dimensional patent evaluation method based on a large language model, characterized by: include: Extract keywords from the patents in the pre-built patent raw data set, perform clustering based on the extracted keywords, and obtain several initial technology clusters; For each initial technology cluster obtained, optimize it using a pre-trained large language model to obtain an optimized technology cluster; Based on the keywords extracted from the patents and combined with the optimized technology clusters, the technology cluster to which the current patent belongs is determined, and the patents belonging to the same technology cluster are regarded as a patent cluster to realize the construction of patent clusters; The patent cluster network is constructed by taking the patents in the patent cluster as nodes and the similarities between patents as edges. Based on the constructed patent cluster network, the importance score of each patent is calculated. Based on the importance score of the patent, combined with the number of applicants, the number of patents, the number of inventors of the cluster to which the patent belongs, and the geographical distribution of the patent, the evaluation results of the patent are obtained.

2. The patent multi-dimensional evaluation method based on a large language model as claimed in claim 1, characterized in that: The method extracts keywords from the patents in the pre-constructed original patent data set, performs clustering based on the extracted keywords, and obtains several initial technology clusters, specifically: divides the patents in the original patent data set into several categories based on the patent IPC classification number; extracts keywords from the patents in each category; and obtains clustering results in each category through cluster analysis based on the extracted keywords; Based on the clustering results obtained under each category, several initial technology clusters are obtained through secondary clustering.

3. The patent multi-dimensional evaluation method based on a large language model as claimed in claim 2 is characterized in that: The method of obtaining clustering results under each category through cluster analysis based on the extracted keywords is as follows: clustering the keywords under the current category based on a preset clustering algorithm to obtain a first clustering set and a number of isolated keywords that are not clustered; clustering the isolated keywords and the keywords in the largest cluster in the first clustering set using a preset clustering algorithm to obtain a second clustering set; The first cluster set and the second cluster set are merged to obtain a clustering result.

4. The patent multi-dimensional evaluation method based on a large language model as claimed in claim 3 is characterized in that: Based on the clustering results obtained under each category, several initial technology clusters are obtained through secondary clustering. Specifically, based on the clustering results obtained under each category, the average vector representation of each member in each cluster is used as the feature representation of the cluster. Based on the feature representation of each cluster, a preset clustering algorithm is used to perform secondary clustering, and based on the obtained clustering results, several technology clusters corresponding to the entire patent original data set are obtained.

5. The patent multi-dimensional evaluation method based on a large language model as claimed in claim 1, characterized in that: Based on the constructed patent cluster network, the importance score of each patent is calculated. Specifically, the PageRank algorithm is used to calculate the PageRank value of each node in the patent cluster network, and the PageRank value is used as the importance score of the patent.

6. The patent multi-dimensional evaluation method based on a large language model as claimed in claim 5 is characterized in that: The calculation of the importance score is specifically based on the following formula: ； in, is the PageRank value of patent v, d is the damping coefficient, N is the total number of patents in the cluster, u, v represents a patent in the cluster, is the set of all patents whose cosine similarity with patent v is higher than the set threshold, is the number of cosine similarities between patent v and other patents that are above the threshold.

7. The patent multi-dimensional evaluation method based on a large language model as claimed in claim 1, characterized in that: The similarity between the patents is calculated based on the similarity of two patent feature vectors, wherein the patent feature vector is obtained by weighted summation of several keyword vectors extracted from the patents, and the weight of each keyword is determined based on the TF-IDF algorithm.

8. The patent multi-dimensional evaluation method based on a large language model as claimed in claim 1, characterized in that: Determining the technology cluster to which the current patent belongs is specifically: based on the keywords corresponding to the patent, querying the technology clusters to which these keywords belong, if the technology cluster is found, the current patent belongs to the technology cluster; If the corresponding technology cluster is not found, the similarity between the patent keywords and each technology cluster is calculated, and the technology cluster to which the current patent belongs is determined based on the similarity results.

9. The patent multi-dimensional evaluation method based on a large language model as claimed in claim 1, characterized in that: The evaluation results of the patents are obtained by combining the importance score of the patents with the number of applicants, the number of patents, the number of inventors and the geographical distribution of the patents in the cluster to which the patents belong, as shown below: ； ； ； ； in, is the logarithmic function with base e, , , , and They are the weights corresponding to the importance score of patents, the number of patents, the number of applicants, the number of inventors, and the geographical distribution of patents; , , , and They respectively represent the patent importance score, the number of patents, the number of applicants, the number of inventors, and the geographical distribution of patents.

10. A patent multi-dimensional evaluation system based on a large language model, characterized by: include: An initial technology cluster construction unit, which is used to extract keywords from patents in a pre-constructed original patent data set, perform clustering processing based on the extracted keywords, and obtain a number of initial technology clusters; A technology cluster optimization unit, which is used to optimize each obtained initial technology cluster using a pre-trained large language model to obtain an optimized technology cluster; Patent cluster construction unit, which is used to determine the technology cluster to which the current patent belongs based on the keywords extracted from the patent and the optimized technology cluster, and to regard the patents belonging to the same technology cluster as a patent cluster to realize the construction of patent cluster; The importance scoring unit is used to construct a patent cluster network using the patents in the patent cluster as nodes and the similarities between patents as edges; based on the constructed patent cluster network, the importance score of each patent is calculated; The patent evaluation unit is used to obtain the evaluation results of the patent based on the importance score of the patent, combined with the number of applicants, the number of patents, the number of inventors and the geographical distribution of the patent in the cluster to which the patent belongs.

Images