Technical route generation method based on autonomous learning

Abstract

The invention relates to a technical route generation method based on autonomous learning. The technical route generation method comprises the steps of automatic document retrieval and content extraction; recursive clustering is carried out based on an optimized vMF-MM algorithm, a structured semantic hierarchical tree is generated, dimension reduction processing is carried out on a high-dimensional vector, multiple view angles of LLM are fused, dynamic soft clustering is carried out by using cosine similarity and adaptive concentration, and the clustering stability under small samples is remarkably improved; a leaf node cluster most matched with the user demand is found in the hierarchical tree; selecting the literature with the highest correlation with the demand from the matched leaf node cluster as a reference, and introducing an MMR diversity reordering mechanism to iteratively screen other candidate items; the problem of recommendation result homogenization is solved, and a technical route map which not only meets user requirements but also has diversity and complementation characteristics is generated by utilizing a scheme screened out through LLM comparative analysis.

Description

Technical Field

[0001] This invention relates to the field of automated research tools technology, and more specifically to a method for generating technical routes based on autonomous learning. Background Technology

[0002] At present, mainstream literature analysis technology has evolved from traditional keyword matching to intelligent analysis based on large language model (LLM). Existing related technologies are mainly focused on automated report generation. Existing technologies use APIs to search and filter literature, use large models to generate the topics of each review article and extract relevant paragraphs for integration and polishing; or analyze review literature, use LLM to extract the main content, and construct mind maps. Although these methods effectively improve the efficiency of reading and organizing single or multiple articles, they focus more on the static extraction and visualization of existing information and lack the ability to deeply reason about the feasibility of technology. Although the above-mentioned existing technologies have achieved automation of technology research to a certain extent, they still have the following problems when facing the technology selection needs of high-dimensional semantic space and complex scenarios: (1) Poor clustering effect of high-dimensional small sample: In specific technical fields, there are often very few literature samples retrieved, but the text vector dimension is extremely high (such as 3072 dimensions). In this scenario, existing methods are difficult to achieve effective clustering, resulting in blurred boundaries of clustering results and failure to form a technical structure tree with clear meaning. Furthermore, traditional hard clustering methods based on K-means tend to forcibly break down a broad technical method or ignore some sparse new technologies. Most traditional clustering algorithms are based on Euclidean distance calculation, while the cosine similarity used in this invention is more suitable for text vector clustering. (2) Lack of interpretable technical route reasoning mechanism: Existing technologies can only statically display relevant research directions and corresponding literature clustering results, but cannot perform dynamic technical route reasoning based on the user's specific natural language constraints, and cannot generate accurate technical route matching when facing complex engineering needs. Most existing methods adopt traditional hard clustering methods. Traditional top-down classification tree search often uses greedy algorithms. Once a classification error occurs at the root node or intermediate level, the subsequent path will completely deviate from the correct direction, lacking a soft routing mechanism that maps user scenarios to technical levels. In existing research recommendations based on large models, redundant recommendations are prone to occur. There is a lack of comparative analysis between different technical solutions, and it is impossible to provide technical route recommendations that meet the needs and are complete. Therefore, a technical route generation method based on autonomous learning is urgently needed to solve the above problems. Summary of the Invention

[0003] The purpose of this invention is to provide a method for generating technical routes based on autonomous learning, which can automatically execute the complete academic research process and generate professional and feasible technical routes; it solves the problems of inaccurate literature clustering and lack of dynamic technical route generation under small sample conditions, and realizes end-to-end reasoning from natural language scene description to dynamic technical route recommendation.

[0004] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: a method for generating technical routes based on autonomous learning, comprising the following steps: S1. Perform literature retrieval based on user input requirements, and extract and summarize the content of the retrieval list results; S2. Based on the extracted and summarized content, technical dimensions are obtained through LLM, abstracts of documents are extracted from specific perspectives, and transformed into vectors through the Embedding model; The vMF-MM algorithm is optimized by introducing prior cluster classification weights and adaptive concentration parameters; recursive clustering is performed based on the optimized vMF-MM algorithm to generate a structured semantic hierarchical tree; S3. Perform scenario analysis and semantic alignment on the input requirements, and calculate the vMF posterior probability of the query vector belonging to each node based on the optimized vMF-MM algorithm. Use the bundle search mechanism to explore multiple paths and find the leaf node cluster that best matches the user's requirements in the clustering hierarchy tree. S4. Select the literature with the highest relevance to the needs from the matched leaf node clusters as the benchmark, introduce the MMR diversity reordering mechanism to iteratively screen other candidate options; use LLM comparative analysis to select the schemes and generate the final technical route.

[0005] Preferably, in S1, the user's input request is converted into search language, and a list of documents is obtained after the search. The extraction and summarization of the document list results includes: downloading the PDFs from the document list, extracting the text content of the PDFs using a PDF parsing tool, filling the PDF text content into prompts, generating a preliminary summary of the detailed content of each document by LLM, and outputting the structured content of each document, including title, abstract, research overview, main contributions, research methods, experiments, results, and conclusions.

[0006] Preferably, in S2, the vector is further preprocessed, including: reducing the dimension of the vector by PCA, retaining the dimension with a cumulative variance contribution rate ≥90%, and performing L2 normalization on the dimension-reduced vector to obtain the preprocessed vector.

[0007] Preferably, in S2, labels and descriptions for K technology clusters are generated using LLM, the label descriptions are vectorized using an embedding model, and the cluster centers and the centers generated by LLM are merged: ; in, As the center of K-Means cluster, The center generated for LLM, This is a parameter related to the level of trust.

[0008] Preferably, prior cluster classification weights and adaptive concentration parameters are introduced to optimize the vMF-MM algorithm, specifically as follows: Assuming that the proportion of each cluster and the concentration within each cluster are equal, a simplified vMF model is obtained; Preliminary clustering results of K clusters are obtained by calculating based on the simplified vMF model; The proportion of samples in each cluster in the preliminary clustering results is used as the prior weight. ; The cluster's concentration parameter is calculated using the mean cosine similarity within the cluster. ; ; ; in, Let K represent the number of samples in the i-th cluster, and K represent the number of categories into which the data samples are divided. Indicates minimum concentration. Indicates the maximum concentration. Let e ​​be the normalized vector and the cluster direction mean vector. cosine similarity, This represents the average cosine similarity within the i-th cluster. The range is between [0,1]; A parameter representing the cluster's concentration. and These represent the minimum and maximum draw similarities for each cluster, respectively. Based on prior weights and concentration parameters The optimized vMF-MM algorithm is obtained, and the vMF posterior probability calculation based on the optimized vMF-MM algorithm is expressed as follows: ; in, This represents the weight of the j-th cluster among the K clusters. Let represent the concentration parameter of the j-th cluster, and .

[0009] Preferably, during recursive clustering, a hierarchical adaptive strategy is introduced for nodes at different levels, including: When the number of samples in a child node cluster is greater than the preset value N1, clustering is performed using the optimized vMF-MM algorithm. When the sample size within a child node cluster is less than the preset value N1, a fixed concentration parameter k is used, and the clustering confidence parameter in the cluster centers is integrated. The value is 0, and the partitioning is based on the semantics of the child nodes generated by LLM.

[0010] Preferably, the constraints of the recursive clustering are: when the number of nodes in a cluster is less than 5, the node splitting stops; when the average cosine similarity within a cluster is greater than 0.95, no splitting is required.

[0011] Preferably, after obtaining the vMF posterior probability of each node, the method further includes determining the optimal clustering combination through a dynamic search mechanism. The dynamic search mechanism is to maximize the cumulative probability of all documents being assigned to candidate clusters, with the objective function being: max ,in , =1 indicates that document e is selected and assigned to cluster i, and each document can only be assigned to a specific cluster in the end.

[0012] Preferably, in S3, after determining the query vector, a layer-by-layer path filtering is performed based on the optimized vMF-MM algorithm. The vMF posterior probability scores of all nodes are calculated, and the vMF probabilities of each layer are multiplied cumulatively. The top N nodes with probabilities greater than the threshold are retained to enter the next layer. The total score of the corresponding cluster combination is updated. The scores of all cluster combinations are monitored in real time. If the current cumulative score of a cluster combination is lower than the average score of all cluster combinations, it is directly removed. When a leaf node is reached or the cumulative confidence score is lower than the threshold, the recursion stops. The leaf node cluster that best matches the scores of all cluster combinations is locked, and the cluster contains a set of corresponding documents.

[0013] Preferably, in S4, based on the most matching leaf node cluster, the cosine similarity between each document in the leaf node cluster and the input requirement is calculated, and the document with the highest cosine similarity is selected as the first recommendation; the results are then stored in the result set. ; Using the MMR diversity re-ranking mechanism, subsequent recommendations are cyclically selected from the remaining literature, and the results for each candidate literature are calculated. MMR score: ; in, The relevance score is calculated using cosine similarity. Represents the query vector. This is a redundancy penalty term. These are the candidate documents currently being evaluated. For the selected literature, This is an adjustment factor, with a value range of 0 to 1.

[0014] Beneficial effects: (1) This invention optimizes the vMF-MM algorithm by introducing prior cluster classification weights and adaptive concentration parameters, which solves the limitations of traditional clustering methods in clustering small samples and high-dimensional text vectors. It performs dimensionality reduction on high-dimensional vectors, integrates the multiple perspectives of LLM, and uses cosine similarity and adaptive concentration to perform dynamic soft clustering, which significantly improves the clustering stability under small samples.

[0015] (2) This invention solves the problem of how to select the optimal clustering combination from different perspectives to construct the classification nodes at the current level; first, different classification combinations are tested by literature data at the same time, and the schemes with low scores are continuously eliminated in the process of data processing, and the one that is left is the technical route that can most accurately summarize the literature in this direction.

[0016] (3) This invention uses soft clustering and bundle search mechanism to accumulate probability scores of nodes at all levels and recommend technical paths in the global technical dimension; through the multi-class probability output of soft classification, the subsequent dynamic search algorithm can weigh the confidence under different perspectives, thereby finding the globally optimal classification structure; in the process of technical path matching, it no longer only retains the Top-1 root node path, but accumulates the probability scores of nodes at all levels to determine the final leaf node; combined with the bundle search mechanism, the system can handle fuzzy and complex business requirements and has stronger fault tolerance and exploratory capabilities.

[0017] (4) This invention solves the problem of homogeneity of recommendation results by reordering leaf nodes by integrating the MMR algorithm. The technical route not only meets the user requirements, but also explains the decision-making trade-off analysis between different technical solutions through LLM, providing users with in-depth technical selection reference. Attached Figure Description

[0018] Figure 1 This is a flowchart of the technical route generation method based on autonomous learning of the present invention.

[0019] Figure 2 A framework flowchart is generated for the technical route of this invention. Detailed Implementation

[0020] To make the objectives and advantages of this invention clearer, the invention will be specifically described below with reference to embodiments. It should be understood that the following text is merely used to describe one or more specific embodiments of the invention and does not strictly limit the scope of protection specifically claimed by the invention.

[0021] Example: Reference Figures 1-2 As shown, a method for generating technical routes based on autonomous learning includes the following steps: S1. Perform literature retrieval based on user input requirements, and extract and summarize the content of the retrieval list results; This system translates user-input research content into search language; specifically, it uses LLM to generate search keywords and classify journals, generating corresponding search statements; it uses the arXiv API for automatic relevance retrieval, obtaining a large list of documents containing information such as title, abstract, publication date, and PDF; it allows setting search criteria, such as time constraints (last year), source constraints, and the number of results to return. The returned results support multiple sorting methods; arXiv is suitable for English searches, while Chinese searches can be performed using OpenALEX. Next, the PDFs in the search results are downloaded, their content extracted, and summarized sequentially, outputting a JSON structure tree for each article. A PDF parsing tool (PyMuPDF) is used to extract the text content, filling the prompts with the PDF text. An LLM tool then generates a preliminary summary of each article's detailed content and outputs the structured content of each document, including title, abstract, research overview, main contributions, research methods, experiments and results, and conclusions, thus achieving text extraction for each document. S2. To build a knowledge base for the retrieved content, a large number of unstructured documents need to be transformed into a semantic hierarchical structure. Technical dimensions are obtained through LLM, and abstracts of documents from specific perspectives are extracted and transformed into vectors through the Embedding model. Prior cluster classification weights and adaptive concentration parameters are introduced to optimize the vMF-MM algorithm. Recursive clustering is performed based on the optimized vMF-MM algorithm to generate a structured semantic hierarchical tree. Taking the clustering of first-level nodes as an example, the following steps are required: generating the current level dimension labels by combining LLM; generating summaries for each document from each dimension perspective; text vector processing; and using the optimized vMF-MM method for clustering. When determining the next level node labels, the LLM generates fine-grained labels for each research direction based on the clustering results of the previous level. It provides the LLM with the current node's topic path (previous node label) and the literature associated with the current node. Through Prompt constraints, the LLM generates the core dimensions for this level, including the dimension name and a brief description. These generated dimensions guide the generation of abstracts for the literature from that perspective, supporting subsequent clustering. Text vector processing transforms the dimension label descriptions and corresponding literature content descriptions into vectors using an Embedding model.

[0022] During the clustering process, the optimized vMF-MM algorithm needs to be used sequentially to calculate the probability of each perspective at that level. The vMF-MM calculation result yields a probability distribution. , representing the text vector of a document from viewpoint a. The probability of belonging to cluster i; the role of vMF-MM is soft classification. By outputting multi-class probabilities, the subsequent dynamic search algorithm can weigh the confidence from different perspectives, thereby finding the globally optimal classification result.

[0023] The specific implementation of recursive clustering based on the optimized vMF-MM algorithm to generate a structured semantic hierarchical tree is as follows: S21. Utilize LLM to generate core dimension labels (semantic anchors) at each level. Combine LLM prior knowledge and statistical methods to generate labels and descriptions for K technology clusters through LLM. Use an embedding model to vectorize the label descriptions, and fuse the cluster centers and the centers generated by LLM into vectors: ; in, As the center of K-Means cluster, The center generated for LLM, For trust level parameters; In one embodiment, the vectors are preprocessed: For a specific perspective 'a', LLM generates a summary for document 'd' based on that perspective. The text-embedding-v4 model is used to encode the document summary, resulting in a D1-dimensional vector (API parameters support active dimensionality reduction; the default vector dimension is 1024, with a minimum supported dimension of 64 and a maximum of 3072). This set of document vectors is then reduced to D2 dimensions (less than 50, discarding dimensions with less information (i.e., smaller variance)) using PCA. The dimensionality-reduced vectors are then subjected to L2 normalization to obtain the preprocessed vectors. .

[0024] S22. Assuming that the proportion of each cluster and the concentration within each cluster are equal, the simplified vMF model is obtained, including: S221, regarding the posterior probability of vMF Initialize the parameters in the formula: ; ; Here, the numerator represents the weighted probability of belonging to the i-th cluster, and the denominator represents the sum of probabilities of belonging to all K clusters. It is the mixing coefficient of the i-th cluster (prior probability, i.e. the original proportion of this cluster in the overall data). Let be the probability density of vector e falling into the i-th cluster, representing the likelihood that the document belongs to that cluster in the semantic direction; ,in, Let e ​​be the normalized vector and the cluster direction mean vector. The cosine similarity (the larger the value, the closer the document is to the center of the cluster). This is a concentration parameter (the larger the value, the more compact the structure). This is a normalization constant, calculated from the dimension D and the concentration k; S222. In the case of a small sample, two assumptions are made: it is assumed that the probability of all clusters occurring is equal (i.e., = =1 / K); assuming that the concentration of each cluster is equal (i.e., = =k); then the simplified vMF model can be expressed as: ; in, The concentration coefficient represents the degree of semantic matching between the text vector and the cluster. S23. Based on the simplified vMF model, calculate the preliminary clustering results of K clusters; S24. Introducing Prior Weights The proportion of samples from each cluster in the first round of results is used as the statistical measure. Substitute the values ​​into the following formula to calculate the posterior probability again: ; S25. Replace the concentration parameter k in the standard vMF with the intra-cluster average similarity; calculate the average cosine similarity for each cluster i. And perform a linear mapping on k within a manually set concentration range; ; ; in, Let be the number of samples in the i-th cluster. The range is between [0,1], and the closer it is to 1, the denser the samples within the cluster; the cluster concentration is mapped to [ , Within the range (e.g., [10, 50]); and These represent the minimum and maximum draw similarities for each cluster, respectively.

[0025] S26, Based on Prior Weights and concentration parameters The optimized vMF-MM algorithm is obtained, and the vMF posterior probability calculation based on the optimized vMF-MM algorithm is expressed as follows: ; For a given document vector e and cluster i from perspective a, vMF-MM outputs the posterior probability that the document belongs to the cluster. This probability value is based on the vector e and the cluster center. The cosine similarity is calculated and satisfies the normalization constraint, meaning that for the same document, the sum of the probabilities of it belonging to all K candidate clusters is 1; that is... ; In one embodiment, during recursive clustering, a hierarchical adaptive strategy is introduced for nodes at different levels, including: When the number of samples in a child node cluster is greater than the preset value N1, clustering is performed using an optimized vMF-MM algorithm. When the sample size within a child node cluster is less than the preset value N1, a fixed concentration parameter k is used, and the clustering confidence parameter in the cluster centers is integrated. The clustering is performed based on the semantics of child nodes generated by LLM, with the initial value set to 0. During the dynamic generation of child node centers, since it's impossible to predefine the cluster centers for each sub-technical direction, a dynamic generation based on LLM is used. For example, a prompt might be: "This is a summary of technical literature on {parent node name}. Please read this content, summarize the {K} most significant sub-directions in this field, and return the names and brief descriptions of these sub-directions." Then, the K sub-direction names returned by LLM are vectorized and used as the cluster center vector for the current level. ; Before recursive clustering, the following conditions must be determined: Stop splitting nodes when the number of nodes in a cluster is less than 5; stop splitting nodes when the average cosine similarity within the cluster is... A similarity score >0.95 indicates extreme semantic similarity, requiring no further splitting. The final leaf nodes contain detailed information about each document. S27. Posterior probability of vMF at each node Then, the algorithm determines the optimal clustering combination through a dynamic search mechanism, constructing the search space S; this process is achieved by maximizing the cumulative probability that all documents are assigned to candidate clusters, with the objective function being: max ,in , =1 indicates that vector e is selected and assigned to cluster i; the optimal cluster is selected from all candidate clusters generated from all perspectives to form the final classification scheme; the selection results must meet the uniqueness constraint, that is, each document can only be assigned to a specific cluster in the end, so as to ensure the mutual exclusivity of the classification results and the maximization of overall credibility.

[0026] S3. Perform scenario analysis and semantic alignment on the input requirements, and calculate the vMF posterior probability of the query vector belonging to each node based on the optimized vMF-MM algorithm. Use the bundle search mechanism to explore multiple paths and find the leaf node cluster that best matches the user's requirements in the clustering hierarchy tree. S31. First, the user's natural language input is converted into a query vector that the system can understand. LLM is used to decompose the scenario into constraints corresponding to the knowledge base perspective using Prompt; interactive semantic alignment is adopted, which calculates the similarity variance between the initial query vector and the cluster centers of the first-level nodes. If the variance is too large, it indicates that the query direction is ambiguous, and the user should be asked to correct the query vector direction. A normalized query vector is output to ensure semantic alignment. S32. Determining the query vector Subsequently, the system utilizes the vMF-MM model constructed during the clustering phase to perform layer-by-layer path filtering, calculating the score of the query vector belonging to each node. Employing a multi-path exploration approach, it does not simply select the parent node with the highest probability, but instead calculates the matching probability scores of all nodes, multiplies them by the vMF probability of each layer, and retains the top N paths with probabilities greater than a threshold, thus avoiding missed recalls due to ambiguity in upper-layer classification. The path inference output of this step is not a single node, but a set of leaf nodes with confidence distributions. S33. To identify the optimal technology cluster in the semantic hierarchy tree, a bundle search and dynamic pruning strategy were employed. All search spaces S have been constructed during the clustering phase described above, where each combination s consists of K semantic vectors generated during the clustering phase. It will be composed of, and will use the prior weights of each level. and adaptive concentration Calculate the vMF probability and the semantic matching incremental score. Employ a beam search mechanism to retain the top N candidate nodes with probabilities greater than a threshold to proceed to the next layer, and update the total score of combination s.

[0027] During the streaming process, the scores of all combinations are monitored in real time. If the current cumulative score of a combination s is significantly lower than the average score of all combinations, it is directly removed from the search space S. When a leaf node is reached or the cumulative confidence score is lower than the threshold (e.g., 0.45), the recursion stops. Finally, the best technology cluster is locked based on the scores of all combinations, and the cluster contains a set of corresponding literature. S4. Select the literature with the highest relevance to the needs from the matched leaf node clusters as the benchmark, introduce the MMR diversity reordering mechanism to iteratively screen other candidate options; use LLM comparative analysis to select the schemes and generate the final technical route.

[0028] S41. First, obtain the parsed query vector. And N document vectors recommended by the technology path; where the document vector is the vector of vMF-recommended documents from the perspective of this technology direction.

[0029] S42. Next, calculate the vector e of each recommended document and the query vector. Based on the cosine similarity, the most relevant paper is selected as the first recommendation and stored in the results set. .

[0030] S43. Using the MMR diversity re-ranking mechanism, subsequent recommendations are selected cyclically from the remaining literature, and the results for each candidate literature are calculated. MMR score: ; in, The relevance score is calculated using cosine similarity. For redundancy penalty term (if With literature collection (A higher penalty score is given if the content is similar to a certain article in the text). These are the candidate documents currently being evaluated. These are already selected documents; this is to reduce redundant recommendations. This is an adjustment factor, with a value range of 0 to 1. The smaller the value, the more the system tends to recommend diverse technologies; (Assuming document B, first calculate its similarity to the user's needs as 0.9, then calculate the relevance between B and the selected optimal document A as 0.4, assuming...) =0.5, B's MMR score is 0.5*0.9-0.5*0.4=0.25; Suppose another document C has a similarity of 0.95 with the demand and 0.98 with the recommendation A, but its MMR score (0.5*0.95-0.5*0.98=-0.015) is lower than B's. S44. After selecting the Top-K diverse literature based on MMR scores, the discrete literature needs to be transformed into logical technical solutions. First, deep comparative reasoning is performed using a large language model, guided by specific prompts under the constraints of the user's specific business scenario. The system analyzes the advantages, disadvantages, and complementary relationships of K technical solutions to generate explainable recommendation reasons. Subsequently, the system executes an automated mind map construction process, using the nodes at each level determined by the previously established vMF probabilistic inference as a framework to structure the recommendation set. The references in the document are used as leaf nodes. Using tools such as XMind SDK, the pre-extracted JSON content structure tree of each reference is embedded under the corresponding topic tags, ultimately synthesizing a technology roadmap mind map. Based on the generated complete mind map, a comprehensive feasibility analysis is generated using a large model. With the structured mind map as context, the content generated by LLM is more logically rigorous and richer in detail than content generated directly.

[0031] The embodiments of the present invention have been described in detail above with reference to the examples. However, the present invention is not limited to the above embodiments. For those skilled in the art, after learning the contents described in the present invention, several equivalent changes and substitutions can be made without departing from the principle of the present invention. These equivalent changes and substitutions should also be considered to fall within the protection scope of the present invention.

Claims

1. A method for generating technical routes based on autonomous learning, characterized in that: Includes the following steps: S1. Perform literature retrieval based on user input requirements, and extract and summarize the content of the retrieval list results; S2. Based on the extracted and summarized content, technical dimensions are obtained through LLM, abstracts of documents are extracted from specific perspectives, and transformed into vectors through the Embedding model; The vMF-MM algorithm is optimized by introducing prior cluster classification weights and adaptive concentration parameters; recursive clustering is performed based on the optimized vMF-MM algorithm to generate a structured semantic hierarchical tree; S3. Perform scenario analysis and semantic alignment on the input requirements, and calculate the vMF posterior probability of the query vector belonging to each node based on the optimized vMF-MM algorithm. Use the bundle search mechanism to explore multiple paths and find the leaf node cluster that best matches the user's requirements in the clustering hierarchy tree. S4. Select the literature with the highest relevance to the demand from the matched leaf node clusters as the benchmark, and introduce the MMR diversity reordering mechanism to iteratively screen other candidates. The selected solutions are analyzed using LLM to generate the final technical roadmap.

2. The method for generating a technical route based on autonomous learning according to claim 1, characterized in that: In S1, the user's input request is translated into search language, and a list of literature is obtained after the search. The content of the literature list results is extracted and summarized, including: downloading PDF files from the PDF address in the literature list, extracting the text content of the PDF using a PDF parsing tool, filling the PDF text content into prompt words, generating a preliminary summary of the detailed content of each literature by LLM, and outputting the structured content of each literature, including title, abstract, research overview, main contributions, research methods, experiments, results and conclusions.

3. The method for generating a technical route based on autonomous learning according to claim 1, characterized in that: S2 also includes preprocessing the vector, including: reducing the dimension of the vector by PCA, retaining the dimension with a cumulative variance contribution rate ≥90%, and performing L2 normalization on the dimension-reduced vector to obtain the preprocessed vector.

4. The method for generating a technical route based on autonomous learning according to claim 3, characterized in that: In S2, labels and descriptions for K technology clusters are generated using LLM. The embedding model is used to vectorize the label descriptions, and the cluster centers and the centers generated by LLM are merged. ； in, As the center of K-Means cluster, The center generated for LLM, This is a parameter related to the level of trust.

5. The method for generating a technical route based on autonomous learning according to claim 4, characterized in that: The vMF-MM algorithm is optimized by introducing prior cluster classification weights and adaptive concentration parameters, specifically as follows: Assuming that the proportion of each cluster and the concentration within each cluster are equal, a simplified vMF model is obtained; Preliminary clustering results of K clusters are obtained by calculating based on the simplified vMF model; The proportion of samples in each cluster in the preliminary clustering results is used as the prior weight. ; The cluster's concentration parameter is calculated using the mean cosine similarity within the cluster. ; ； ； in, Let K represent the number of samples in the i-th cluster, and K represent the number of categories into which the data samples are divided. Indicates minimum concentration. Indicates the maximum concentration. Let e ​​be the normalized vector and the cluster direction mean vector. cosine similarity, This represents the average cosine similarity within the i-th cluster. The range is between [0,1]; A parameter representing the cluster's concentration. and These represent the minimum and maximum draw similarities for each cluster, respectively. Based on prior weights and concentration parameters The optimized vMF-MM algorithm is obtained, and the vMF posterior probability calculation based on the optimized vMF-MM algorithm is expressed as follows: ； in, This represents the weight of the j-th cluster among the K clusters. Let represent the concentration parameter of the j-th cluster, and .

6. The method for generating a technical route based on autonomous learning according to claim 5, characterized in that: When performing recursive clustering, a hierarchical adaptive strategy is introduced for nodes at different levels, including: When the number of samples in a child node cluster is greater than the preset value N1, clustering is performed using the optimized vMF-MM algorithm. When the sample size within a child node cluster is less than the preset value N1, a fixed concentration parameter k is used, and the clustering confidence parameter in the cluster centers is integrated. The value is 0, and the partitioning is based on the semantics of the child nodes generated by LLM.

7. A method for generating a technical route based on autonomous learning according to claim 5 or 6, characterized in that: The constraints of the recursive clustering are: stop splitting nodes when the number of nodes in a cluster is less than 5; and do not split nodes when the average cosine similarity within a cluster is greater than 0.

95.

8. The method for generating a technical route based on autonomous learning according to claim 5, characterized in that: After obtaining the vMF posterior probability of each node, the optimal clustering combination is determined through a dynamic search mechanism. This dynamic search mechanism maximizes the cumulative probability of all documents being assigned to candidate clusters, with the objective function being: max ,in , =1 indicates that document e is selected and assigned to cluster i, and each document can only be assigned to a specific cluster in the end.

9. The method for generating a technical route based on autonomous learning according to claim 8, characterized in that: In S3, after determining the query vector, a layer-by-layer path filtering is performed based on the optimized vMF-MM algorithm. The vMF posterior probability scores of all nodes are calculated, and the vMF probabilities of each layer are multiplied cumulatively. The top N nodes with probabilities greater than the threshold are retained to enter the next layer. The total score of the corresponding cluster combination is updated. The scores of all cluster combinations are monitored in real time. If the current cumulative score of a cluster combination is lower than the average score of all cluster combinations, it is directly removed. When a leaf node is reached or the cumulative confidence score is lower than the threshold, the recursion stops. The leaf node cluster that best matches the scores of all cluster combinations is locked, and the cluster contains a set of corresponding documents.

10. The method for generating a technical route based on autonomous learning according to claim 1, characterized in that: In S4, based on the most matching leaf node cluster, the cosine similarity between each document in the leaf node cluster and the input requirement is calculated, and the document with the highest cosine similarity is selected as the first recommendation; the results are then stored in the result set. ; Using the MMR diversity re-ranking mechanism, subsequent recommendations are cyclically selected from the remaining literature, and the results for each candidate literature are calculated. MMR score: ； in, The relevance score is calculated using cosine similarity. Represents the query vector. This is a redundancy penalty term. These are the candidate documents currently being evaluated. For the selected literature, This is an adjustment factor, with a value range of 0 to 1.

Images