A patent technology route generation method and device and computer equipment

Abstract

The application belongs to the field of natural language processing, and particularly relates to a patent technology route generation method and device and computer equipment; the method comprises obtaining patent text data and patent item information; a patent technology information extraction model is used to generate a patent technology analysis result; a clustering list of core technology keywords, technology efficacy keywords and technology field keywords is divided into a main cluster and a secondary cluster; the similarity between patent technology analysis results is calculated according to a fusion vector of the main cluster and the secondary cluster; a patent technology map is constructed according to the item information of the patent and the similarity between the patent technology analysis results; a specific technology keyword is searched in the patent technology map, and a patent technology route analysis graph is generated; and a patent technology route prediction model is used to generate a patent technology route. The application applies related technologies of a large language model and a time sequence knowledge graph to patent route generation, and achieves the effect of sequencing patent text information.

Description

Technical Field

[0001] This invention belongs to the field of natural language processing, and specifically relates to a method, apparatus and computer equipment for generating a patented technical route. Background Technology

[0002] With the explosive growth in the number of global patents, the mining and analysis of massive amounts of patent data has become increasingly important. Patent technology roadmap analysis is an analytical method aimed at revealing the main technological development paths and key technological nodes within a specific technological field. It involves in-depth research into historical and existing patent data to understand technological evolution and potential technological trends from a more comprehensive and structured perspective.

[0003] One reason for the difficulty in patent technology analysis lies in the reading and comprehension of patent texts. This requires not only an understanding of the commercial market but also a grasp of the patent technology. Existing patent technology roadmap analyses are mostly based on basic information such as the patent applicant and IPC classification number to construct a patent technology roadmap, which then requires further detailed technical analysis of the patent text by humans, consuming a significant amount of manpower. In the field of natural language processing, regarding patent reading comprehension techniques, due to the characteristics of patent texts themselves and the lack of manually annotated datasets, most research still focuses on keyword extraction and simple text generation, with less than ideal results.

[0004] The main difficulties to overcome in reading and understanding patent texts are as follows: Chinese processing requires word segmentation, which leads to semantic segmentation; patent documents contain a large number of technical terms and abbreviations, and the understanding ability of general language models is limited; there is a lack of large-scale, high-quality manually annotated data; and the reading and understanding of patents is highly subjective, and the results can easily be biased due to the personal subjective factors of practitioners.

[0005] In recent years, Large Language Models (LLMs) have made significant progress, surpassing or approaching human-level performance in language understanding and generation tasks. They leverage large-scale pre-training to acquire linguistic knowledge and learn to represent language as vectors, demonstrating high performance across various NLP tasks. Recently released LLM models such as ChatGLM and LLaMA have exhibited expert-level language understanding and generation capabilities in multiple domains. Based on the training process of the RLHF (Large Language Model RLHF), some papers have further proposed the RLAIF method, utilizing the capabilities of large language models to train other language models. This allows researchers to use large language models for patent reading comprehension, extracting high-quality keywords by constructing appropriate cue words to compensate for the lack of manually annotated datasets in this area. This can, to some extent, avoid the subjectivity of manual annotation and significantly reduce the cost of manual annotation.

[0006] In constructing patent technology graphs, most studies struggle to achieve satisfactory results on broad datasets due to the limitations of various NLP tasks in the patent domain. Furthermore, most patent knowledge graph research remains focused on analyzing the similarity of technical phrases in the semantic analysis of patent texts, heavily reliant on the quality of extracted technical phrases. Moreover, the lack of clarity and effectiveness in semantically interpreting patent texts hinders the further application of patent knowledge graphs to other related algorithms. Summary of the Invention

[0007] To address the problems existing in current technologies, this invention proposes a method, apparatus, and computer device for generating patent technology routes. It applies technologies related to large language models and time-series knowledge graphs to patent route generation. Utilizing the large language model, key training methods such as RLHF and RLAIF are employed to train a language model with patent text reading comprehension capabilities. This language model can extract core technologies, technical effects, and keywords related to the technical field from the patent, achieving the effect of sequencing the patent text information. Constructing and visualizing a patent technology graph based on the sequenced patent text information clearly describes the main technical routes of the patent. Using the patent technology graph, a corresponding time-series dataset is obtained. This dataset is then used to train a time-series prediction model to achieve the effects of predicting patent technology hotspots and mining patent points in cross-technical fields.

[0008] In a first aspect, the present invention provides a method for generating a patented technology route, the method comprising:

[0009] Obtain patent text data and patent copyright information;

[0010] A pre-trained patent technology information extraction model is used to process patent text data to generate patent technology analysis results; the patent technology analysis results include a cluster list of core technology keywords, technical efficacy keywords, and technical field keywords of the patent.

[0011] The cluster lists of core technology keywords, technical efficacy keywords, and technical field keywords are sorted according to cluster size, and primary clusters and secondary clusters are divided.

[0012] The similarity between the patent technology analysis results is calculated based on the fusion vector of the primary and secondary clusters.

[0013] A patent technology map is constructed based on the similarity between the copyright information of the patents and the results of patent technology analysis.

[0014] Search for specific technical keywords in the patent technology map to generate a patent technology roadmap analysis diagram;

[0015] A pre-trained patent technology route prediction model is used to process the patent technology route analysis diagram to generate patent technology routes.

[0016] In a second aspect, the present invention also provides an apparatus for generating a patented technology route, the apparatus comprising:

[0017] The data acquisition module is used to acquire patent text data and patent copyright information;

[0018] The data generation module is used to process patent text data using a pre-trained patent technology information extraction model to generate patent technology analysis results. The patent technology analysis results include a core technology cluster list, a technology efficacy cluster list, and a technology field keyword cluster list for the patent.

[0019] The patent technology graph construction module is used to sort the core technology cluster list, the technical efficacy cluster list, and the technical field keyword cluster list according to the cluster size, and divide them into primary and secondary clusters; calculate the similarity between the patent technology analysis results based on the fusion vector of the primary and secondary clusters; and construct a patent technology graph based on the patent copyright information and the similarity between the patent technology analysis results.

[0020] The route analysis module is used to search for specific technical keywords in the patent technology map and generate a patent technology route analysis map;

[0021] The route generation module is used to process the patent technology route analysis diagram using a pre-trained patent technology route prediction model to generate patent technology routes.

[0022] In a third aspect, the present invention also provides a computer device comprising a processor and a memory, the memory storing at least one computer program, the at least one computer program being loaded and executed by the processor to implement the method for generating the patented technology route as described in the first aspect of the present invention.

[0023] The patent technology roadmap analysis method described in this invention has the following advantages:

[0024] 1. By leveraging the semantic understanding capabilities of existing large language generation models, the accuracy and fine-grainedness of patent technology analysis are significantly enhanced. Furthermore, clustering algorithms are used to evaluate the content generated by the model, resulting in an unsupervised and objective acquisition of a patent keyword dataset.

[0025] 2. This invention utilizes the RLAIF method to perform reinforcement learning on a pre-trained language model for patent technology analysis tasks. This allows the invention to achieve performance comparable to large language models on patent keyword extraction tasks (i.e., patent technology information extraction models) with fewer resources and a smaller model, and it is also easy to iterate and update on new data.

[0026] 3. The patent time series map constructed by this invention integrates the core technology, technical efficacy, technical field and other basic information of the patent. The time series dataset constructed on this basis has multiple patent technical features. It can be used to predict patent hotspots and patent technology improvements through time series prediction models, and effectively generate the technical route of the patent. Attached Figure Description

[0027] Figure 1 This is a flowchart of a method for generating a patented technical route according to an embodiment of the present invention;

[0028] Figure 2 This is a structural diagram of a device for generating a patented technical route according to an embodiment of the present invention. Detailed Implementation

[0029] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. The present invention will be described in detail below with reference to specific examples.

[0030] Figure 1 This is a flowchart of a method for generating a patented technical route according to an embodiment of the present invention, such as... Figure 1 As shown, the method includes:

[0031] 101. Obtain patent text data and patent copyright information;

[0032] In this embodiment of the invention, existing information can be obtained from patent databases. These databases can be publicly available patent literature repositories from various countries online, or they can be local patent databases established locally by patent analysts who search publicly available patent literature repositories online. The patent text data can include the patent claims, invention content, and specific embodiments, and may also specifically include the technical field, patent description, key entity information, etc. The patent copyright information includes applicant information, application date information, grant date information, IPC classification number, etc.

[0033] 102. A pre-trained patent technology information extraction model is used to process the patent text data to generate patent technology analysis results. These results include a cluster list of core technology keywords, technical efficacy keywords, and technical field keywords for the patent. In this embodiment, the patent technology information extraction model is primarily used to perform technical analysis on the patent text, extracting the core technology, technical efficacy, and technical field keywords involved in the patent. This embodiment uses the open-source large language model LLaMA as its base model, which is based on a transformer-based decoder-only architecture and uses publicly available high-quality datasets. The dataset for patent text analysis based on the large language model is generated; a reward model (RM) is trained; and reinforcement learning (RL) is employed. This completes the training of a language model for extracting specific patent keywords, achieving results equivalent to the large language model.

[0034] In this embodiment of the invention, it is only necessary to obtain the LLaMA large language model and use it as the initial model. The initial language model (LM) is then trained using datasets obtained from GitHub and Wikipedia. In this embodiment, for generating the dataset for patent text analysis based on the large language model, this invention needs to utilize existing, high-performing large language models (such as GPT-4) to complete the dataset construction. The specific steps are as follows:

[0035] 1) Obtain the patent text training set data, and preprocess the patent text training set data to extract the patent content of the patent text training set data, wherein the patent content includes the invention content and the claims.

[0036] 2) Fill the extracted patent content into a pre-built prompt word template to form prompt words; wherein, the prompt word template is constructed as follows:

[0037] Contents: {Summary of the Invention / Claims}

[0038] Question: {Question}

[0039] Format: {format}

[0040] Based on the above 'content', answer the 'question' using the 'format' provided.

[0041] Then fill in the relevant content into the template to form prompt words. The question can be constructed as "Based on the invention content / claims of this patent, return 1-3 keywords that you think are reasonable for technical effects / technical fields / application fields." The format can be constructed as "{keyword 1, keyword 2, keyword 3}".

[0042] 3) Input the prompt word into the large language model, repeat K times, and after formatting and splitting the answer, obtain a list of K keywords;

[0043] 4) Cluster each group of keywords. First, iterate through all keywords in the K groups of keyword lists, and then obtain the word vector key for each keyword key using a word vector model. vec ;

[0044] 5) Randomly select a set of generated content keywords, and use this set of keywords as the initial cluster. Each keyword key forms its own cluster, and its initial cluster center is set as the word vector key of that keyword. vec ;

[0045] In this embodiment of the invention, a set of generated content keywords can be randomly selected, and this set of keywords can be used as the initial cluster. Each keyword key constitutes its own cluster, and its initial cluster center O = [y1, y2, y3, ..., y N Let ] be the word vector key for this keyword. vec N represents the number of cluster centers.

[0046] 6) Iterate through the other groups of keywords, and for each group of X keywords, obtain the keyword vector key. vec Calculate the Euclidean distance D between the keyword and each cluster center; select the cluster with the closest distance to the keyword; if the closest distance is less than a threshold, add the keyword to the cluster and recalculate the new cluster center O based on the existing Y keywords in the cluster; if the closest distance is greater than the threshold, treat the keyword as a new cluster, and the center of the new cluster is the word vector key of the keyword. vec ;

[0047] The new cluster centers O are recalculated based on the existing Y keywords in the clustering as follows:

[0048]

[0049] X and Y both represent the number of keywords. The values ​​of X and Y can be the same or different. Those skilled in the art can set them according to the actual situation. If the nearest distance is greater than the threshold, the keyword is regarded as a new cluster, and the center of the new cluster is the word vector of the keyword.

[0050] 7) Repeat the above process until all keywords are merged to obtain the final clustering result; select the word whose word vector is closest to the cluster center in each cluster as the keyword; sort the clusters from largest to smallest according to their size, and sort each word in each cluster according to the distance of each word's word vector from the cluster center, obtaining a score for each keyword's cluster based on the sorting; obtain a score for each generated content based on the score of each keyword's cluster; output the prompt word, answer, and score as the keyword score. Specifically, assign a value to each word according to the sorting:

[0051]

[0052] Where N is the total number of keywords, n is the cluster size, and k is the keyword's rank within the cluster. The score for each (prompt, answer) is calculated as follows:

[0053]

[0054] Where m is the number of keywords separated in the answer, and score key(i) For each keyword cluster, a score is assigned, and finally, for each (prompt, answer) cluster, a data Keyword_score in the form (prompt, answer, score) is generated.

[0055] In this embodiment of the invention, for the training of the reward model (RM), the present invention trains the reward model RM to assist in the fine-tuning of the initial language model.

[0056] The training method for the reward model RM is described below:

[0057] ① Use a pre-trained language model with a transformer architecture, such as GPT-2, as the base model for the reward model. Add a linear transformation layer at the end of the base model so that the variable output by the model is mapped to a numerical value, thus forming the reward model.

[0058] ②The RM model can be viewed as a scoring model with input (prompt, answer) and output (score). Its loss function is set as follows:

[0059]

[0060] Where n represents the number of samples, m represents the number of keywords in a single sample, and score i key_score represents the label of the i-th sample. j This represents the score of the j-th keyword in each sample.

[0061] In this embodiment of the invention, the reward model RM obtained in the above embodiments is used to perform reinforcement learning on the LM reference PPO algorithm. The specific steps are as follows:

[0062] ① First, define the policy network of the reinforcement learning model (RL) with LM as the initial state, prompt as the input, and keyword text as the output; define the reward function as the pre-trained RM model.

[0063] ②The loss function of the reinforcement learning model is defined as follows:

[0064]

[0065] Where x represents a question in the dataset, y represents the answer obtained by inputting question x into the RL model, r(x,y) represents the score obtained by inputting question x and answer y into the reward model RM, fm(y) represents whether the format of y meets the requirements, and takes a value of 1 if it meets the requirements, otherwise it takes a value of -1, RL(y|x) represents the probability that question x will obtain the answer y through the RL model, LM(y|x) represents the probability that x will obtain the answer y through the model LM(y|x), and α and β are adjustment coefficients.

[0066] The training of the final RL model is now complete. The pre-set prompt template is used as the input to the RL model, and the trained model is used to generate the same dataset based on the large language model for patent text analysis, so that the technical analysis results of the patent text can be extracted.

[0067] 103. The clustering lists of core technology keywords, technical efficacy keywords, and technical field keywords are sorted according to cluster size to divide them into primary and secondary clusters. In this embodiment of the invention, the obtained patent technology analysis results should be three clustering lists, representing the core technology, technical efficacy, and technical field of the patent, respectively. Each list has a length of N and contains information such as the cluster center vector and cluster size. The vector representation V of the core technology, technical efficacy, and technical field of the patent node is calculated. t as follows:

[0068]

[0069] Where N represents the total number of clusters, V i Let n represent the cluster center of the i-th cluster. i Let n represent the cluster size of the i-th cluster. a This represents the sum of all cluster sizes. Clusters are sorted by size, and the top K clusters are selected as primary clusters, while the remaining NK clusters are secondary clusters.

[0070] 104. Calculate the similarity between the patent technology analysis results based on the fusion vectors of the primary and secondary clusters;

[0071] In this embodiment of the invention, time-series information is connected by directed edges based on the magnitude of the time-series information of two patent nodes, pointing from the earlier node to the later node. To connect the technical analysis results of a patent node A with other patent nodes, the fusion vectors of its primary and secondary clusters within its technical field need to be calculated sequentially, as follows:

[0072]

[0073] Among them, V k V represents the k-th fusion vector of the keyword. k' V represents the k-th primary cluster of the keywords. k” This represents the k-th minor cluster of the keyword, NK represents the number of minor clusters, N is the total number of clusters, and K is the number of major clusters; n a n represents the sum of the sizes of all clusters. ck” V represents k” The cluster size of the cluster is determined. Then, the distance between each principal vector between two patent nodes is calculated, and the distances of vectors less than a threshold are used to form a weighted edge. If such a weighted edge exists between two patent nodes, its weight is the Euclidean distance between the two principal vectors.

[0074] 105. Based on the similarity between the patent copyright information and the patent technology analysis results, a patent technology map is constructed. In this embodiment of the invention, undirected edges are constructed for patent nodes with the same time sequence information in the patent copyright; directed edges are constructed for patent nodes with different time sequence information (e.g., patent grant date) in the patent copyright, and the directed edges are connected according to the size of the time sequence information of the two patent nodes; edge weights are constructed according to the similarity between the patent technology analysis results of the patent nodes, that is, the edge weights of the corresponding directed or undirected edges are used as the edge weights of the similarity exceeding a preset threshold, and the directed or undirected edges of the similarity not exceeding the preset threshold are deleted; a technology map is constructed for the patent nodes according to the directed edges, undirected edges, and edge weights.

[0075] 106. Search for specific technical keywords in the patent technology map to generate a patent technology roadmap analysis diagram;

[0076] In this embodiment of the invention, a specific technical keyword is searched in the patent technology map, and a technical path analysis diagram is generated with the technical keyword as the main branch and related technologies diverging in the form of a tree diagram, with time as the axis.

[0077] 107. The patent technology map is processed using a pre-trained patent technology route prediction model to generate patent technology routes. In this embodiment of the invention, it is necessary to train the patent technology route prediction model, and the steps are as follows:

[0078] A time-series dataset is generated based on the patent technology graph. Starting from any node with the least time-series information, and under a limited number of hops, a breadth-first traversal of the undirected graph is performed to obtain several subgraphs arranged according to the patent grant dates in the graph:

[0079] [G1,G2,...,G n ]

[0080] For example, if the hop count is limited to 5, then the resulting subgraph contains all nodes that are 5 hops away from the target node.

[0081] Each subgraph contains a node feature matrix A and an adjacency matrix B. Node feature matrix A is an N×F matrix, where N is the number of nodes in the subgraph, and F is the vector dimension representing the nodes. The core technology vector, technical field vector, and technical efficacy vector of the patent are concatenated to represent the patent node vector. Adjacency matrix B is an N×N matrix representing the connection between any two nodes in the weighted undirected graph. This sequence is used to train a time-series prediction model, specifically using the Informer time-series prediction model. The patent grant date is used as the global time-series information input. Each subgraph is first transformed into a vector representation of the time-series graph through a pre-designed graph convolutional layer, which can be designed as a GAT layer to represent the subgraph as a vector. This vector serves as the input for a time step and outputs as a vector representation of a node. This vector representation contains information on the core technology, technical field, and technical efficacy of the predicted patent. It can be used to predict the technological development hotspots and improvement points of a patent in a specific technical field, helping companies to proactively plan their patent portfolios or focus on the development direction of new technologies in a particular technical field, thus summarizing the technological development of patents and assisting in patent strategy planning.

[0082] This model can be used to predict further technical improvements to a patent and which technologies can be incorporated to achieve improved technical efficacy. It can also identify patent opportunities in cross-technology areas of the latest patents, thereby assisting in patent strategy planning.

[0083] Figure 2 This is a structural diagram of a device for generating a patented technical route according to an embodiment of the present invention, such as... Figure 2 As shown, the device includes:

[0084] The data acquisition module is used to acquire patent text data and patent copyright information;

[0085] The data generation module is used to process patent text data using a pre-trained patent technology information extraction model to generate patent technology analysis results. The patent technology analysis results include a core technology cluster list, a technology efficacy cluster list, and a technology field keyword cluster list for the patent.

[0086] The patent technology graph construction module is used to sort the core technology cluster list, the technical efficacy cluster list, and the technical field keyword cluster list according to the cluster size, and divide them into primary and secondary clusters; calculate the similarity between the patent technology analysis results based on the fusion vector of the primary and secondary clusters; and construct a patent technology graph based on the patent copyright information and the similarity between the patent technology analysis results.

[0087] The route analysis module is used to search for specific technical keywords in the patent technology map and generate a patent technology route analysis map;

[0088] The route generation module is used to process the patent technology route analysis diagram using a pre-trained patent technology route prediction model to generate patent technology routes.

[0089] In this embodiment of the invention, a computer device is also provided, the computer device including a processor and a memory, the memory storing at least one computer program, the at least one computer program being loaded and executed by the processor to generate a patented technology route as described in this invention.

[0090] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, which may include ROM, RAM, disk, or optical disk, etc.

[0091] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for generating a patented technology route, characterized in that, The method includes: Obtain patent text data and patent copyright information; A pre-trained patent technology information extraction model is used to process patent text data to generate patent technology analysis results; the patent technology analysis results include a cluster list of core technology keywords, technical efficacy keywords, and technical field keywords of the patent. The cluster lists of core technology keywords, technical efficacy keywords, and technical field keywords are sorted according to cluster size, and primary clusters and secondary clusters are divided. The similarity between the patent technology analysis results is calculated based on the fusion vector of the primary and secondary clusters. A patent technology map is constructed based on the similarity between the copyright information of the patents and the results of patent technology analysis. Search for specific technical keywords in the patent technology map to generate a patent technology roadmap analysis diagram; The patent technology route analysis diagram is processed using a pre-trained patent technology route prediction model to generate a patent technology route. The pre-training process of the patent technology information extraction model includes: acquiring patent text training set data; inputting the patent text training set data into a large language model, generating patent text training data using a specific template, wherein the patent text training data includes prompt words, answer responses, and keyword scores; training a reward model using the patent text training data to generate content quality information of the patent training text data; and training a reinforcement learning model using the content quality information to generate patent technology analysis results of the patent training text data. Inputting the patent text training set data into a large language model and generating patent text training data using a specific template includes: 1) Obtain the patent text training set data, preprocess the patent text training set data, and extract the patent content from the patent text training set data, wherein the patent content includes the invention content and the claims. 2) Fill the extracted patent content into the pre-built prompt word template to form prompt words; 3) Input the prompt word into the large language model, repeat K times, and after formatting and splitting the answer, obtain a list of K keywords; 4) Cluster each group of keywords. First, iterate through all keywords in the K groups of keyword lists, and then obtain each keyword using a word vector model. word vectors ; 5) Randomly select a set of generated content keywords, and use this set of keywords as the initial cluster. For each keyword... Each cluster forms its own cluster, with its initial cluster center set as the word vector of the keyword. ; 6) Iterate through the other groups of keywords, and for each of the X keywords in the group, obtain the word vector of each keyword. Calculate the Euclidean distance D between the keyword and each cluster center; select the cluster with the closest distance to the keyword; if the closest distance is less than a threshold, then remove the keyword from the cluster. Join the cluster and recalculate the new cluster centers based on the existing Y keywords in the cluster. If the nearest distance is greater than the threshold, then the keyword will be removed. As a new cluster, the center of the new cluster is the keyword. word vectors ; 7) Repeat the above process until all keywords are merged to obtain the final clustering result; take the word whose word vector is closest to the cluster center in each cluster as the keyword; sort the clusters from largest to smallest according to their size, and sort each word in each cluster according to the distance of each word's word vector from the cluster center, and obtain the score of each keyword's cluster according to the sorting; obtain the score of each generated content according to the score of each keyword's cluster; output the prompt, answer, and score as the keyword score.

2. The method for generating a patent technology route according to claim 1, characterized in that, The score for each keyword's cluster is represented as follows: in, This indicates the score of the cluster to which the keyword belongs. Indicates cluster size, This indicates the ranking of keywords within a cluster. This indicates the total number of keywords.

3. The method for generating a patent technology route according to claim 1, characterized in that, The process of constructing a patent technology map based on the similarity between patent copyright information and patent technology analysis results includes: constructing undirected edges for patent nodes with the same time sequence information in the patent copyright; constructing directed edges for patent nodes with different time sequence information in the patent copyright, and connecting the directed edges according to the magnitude of the time sequence information of the two patent nodes; and constructing edge weights according to the similarity between the patent technology analysis results of the patent nodes, that is, using the edge weights with similarity exceeding a preset threshold as the edge weights of the corresponding directed or undirected edges, and deleting the directed or undirected edges with similarity not exceeding the preset threshold. The patent nodes are used to construct a technology graph based on directed edges, undirected edges, and edge weights.

4. The method for generating a patent technology route according to claim 1, characterized in that, The step of searching for specific technical keywords in the patent technology map and generating a patent technology roadmap analysis diagram includes searching for specific technical keywords in the patent technology map, using time as the axis, and generating a technology roadmap analysis diagram with the technical keyword as the main branch and related technologies diverging in the form of a tree diagram.

5. The method for generating a patent technology route according to claim 1, characterized in that, The process of using a pre-trained patent technology route prediction model to process the patent technology route analysis graph and generate a patent technology route includes generating a time-series sequence dataset based on the patent technology route analysis graph. Starting from any node with the smallest time-series information, and under the premise of limiting the number of hops, a sequence of several subgraphs arranged according to the time-series information in the patent technology route analysis graph is obtained by breadth-first traversal of the undirected graph. Each subgraph contains a node feature matrix A and an adjacency matrix B: Node feature matrix A is an N×F matrix, where N is the number of nodes in the subgraph and F is the vector dimension representing the nodes in the subgraph. The cluster center vectors of the clustering list of core technology keywords, technical efficacy keywords, and technical field keywords are used as the vector representation of the patent nodes. Adjacency matrix B is an N×N matrix representing the connection between every two patent nodes in the weighted undirected graph. The sequence is used to train a time-series prediction model, where the time-series information is used as global time-series information input. Each subgraph is first transformed into a vector representation of the time-series graph through a pre-designed graph convolutional layer, which serves as the input for one time step. The output is a vector representation of a patent node, which includes the core technology, technical field, and technical efficacy information of the predicted patent.

6. A patent technology route generation apparatus, wherein the apparatus applies the patent technology route generation method according to any one of claims 1-5, characterized in that, The device includes: The data acquisition module is used to acquire patent text data and patent copyright information; The data generation module is used to process patent text data using a pre-trained patent technology information extraction model to generate patent technology analysis results. The patent technology analysis results include a core technology cluster list, a technology efficacy cluster list, and a technology field keyword cluster list for the patent. The patent technology graph construction module is used to sort the core technology cluster list, the technical efficacy cluster list, and the technical field keyword cluster list according to the cluster size, and divide them into primary and secondary clusters; calculate the similarity between the patent technology analysis results based on the fusion vector of the primary and secondary clusters; and construct a patent technology graph based on the patent copyright information and the similarity between the patent technology analysis results. The route analysis module is used to search for specific technical keywords in the patent technology map and generate a patent technology route analysis map; The route generation module is used to process the patent technology route analysis diagram using a pre-trained patent technology route prediction model to generate patent technology routes.

7. A computer device, characterized in that, The computer device includes a processor and a memory, the memory storing at least one computer program, which is loaded and executed by the processor to implement the method for generating the patented technology route as described in any one of claims 1 to 5.

Images