Intelligent patent technology brief report pushing method and system based on user portrait and large model

Abstract

The invention provides a patent technology brief report intelligent pushing method and system based on a user portrait and a large model, and the method comprises the steps: collecting user dominant identity information and hidden interaction behavior data, and constructing a user portrait vector containing role labels and technical interest distribution based on time distance and interaction depth dynamic weighting; the method comprises the following steps: disassembling patent literatures to be pushed by utilizing a domain knowledge fine-tuning large language model, extracting triple and claim core features of'technical problems, technical means and technical effects', and mapping the triple and claim core features into patent value vectors; and calculating the semantic similarity of the two patents, and determining a target patent by combining technical novelty and competitive risk index weighted correction. And taking the triple as a generation substrate, calling a role attention mask mechanism, and generating and pushing a personalized technical brief report adaptive to a user role through differentiated recombination and rewriting of a large language model. And collecting user feedback data, and dynamically adjusting weighted parameters by using a reinforcement learning algorithm to form closed-loop optimization.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to a patented intelligent push method and system for technology briefings based on user profiles and large models. Background Technology

[0002] Patent information has become a key resource for enterprise technological innovation and strategic decision-making. Existing patent information services mostly rely on keywords, IPC classification numbers, or simple collaborative filtering algorithms for patent recommendations. These traditional methods often fail to capture the dynamic and evolving technical needs of users, resulting in low accuracy and high noise levels in the recommendations. In existing technical solutions, user interest characteristics are typically treated as static tags, ignoring the decay of user interest over time and the differences in interest depth represented by different interactive behaviors, making it difficult to construct accurate dynamic profiles.

[0003] In addition, traditional patent push content usually directly displays patent abstracts or full texts, which are obscure, lengthy and monotonous, and cannot distinguish the different needs of different user roles for information focus. As a result, the cognitive cost for users to obtain key intelligence from massive amounts of information is extremely high.

[0004] Meanwhile, most existing recommendation algorithms are based on similarity matching, which can easily trap users in an "information cocoon." They lack mechanisms for exploring technological novelty and are unable to effectively warn of potential competitor risks based on complex referencing networks. They cannot meet the advanced needs of enterprises for intelligence foresight and risk control in fierce technological competition. Summary of the Invention

[0005] The main objective of this invention is to provide a method and system for intelligently pushing patent technology briefings based on user profiles and large models. This solves the technical problems existing in the patent push technology, such as the static construction of user profiles leading to inaccurate demand matching, the lack of role-specific push content leading to difficulties in information digestion, and the lack of in-depth mining of technological novelty and competitive risks leading to low intelligence value.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by this invention is: a patent technology briefing intelligent push method based on user profiles and large models, comprising: S1. Obtain explicit user identity information and implicit interaction behavior data, and dynamically weight the user's technical interest characteristics based on the time proximity and interaction depth of the user behavior to obtain a user profile vector containing user role tags and technical interest distribution. S2. Using a large language model fine-tuned with domain knowledge, semantically decompose the patent documents to be pushed, extract the structured triple information of "technical problem, technical means, and technical effect" that can represent the essence of the patent technology, as well as the core features of the claims, and map the extracted information into a patent value vector; S3. Calculate the semantic similarity between the user profile vector and the patent value vector, and introduce the technical novelty index and the competition risk index to weight and correct the semantic similarity. Based on the weighted comprehensive score, determine the target patent to be pushed. S4. Using the structured triplet information of the target patent as the generation basis, based on the user role label in the user profile vector, call the corresponding role attention mask mechanism, use the large language model to differentiate and rewrite the generation basis, generate a personalized technical briefing adapted to the user role and push it. S5. Collect user feedback data on the push notifications and dynamically adjust the weight parameters of the weighted correction in S3 using reinforcement learning algorithms.

[0007] In the preferred scheme, the dynamic weighting of user technical interest features in step S1 is implemented based on a time decay and interaction depth weighting algorithm, and the steps are as follows: S11. Preset a basic weight system, setting download and bookmark behaviors as high weight levels and browsing and searching behaviors as low weight levels to determine the basic weight of a single interaction behavior; S12. Obtain the timestamp of the interaction, calculate the time difference between the timestamp and the current time, and use the exponential decay logic to calculate the time decay coefficient corresponding to the time difference, so that the closer the behavior is to the current time, the larger the retention coefficient. S13. Determine the depth adjustment factor based on the interaction depth level, correct the basic weight, and multiply the corrected basic weight by the time decay coefficient to obtain the effective contribution value of a single behavior. S14. Based on the above steps, calculate the final weight of the specific technical topic, using the following formula: ; in, This indicates the final weight of the technical topic. The base weight representing a single interaction behavior. This represents the time difference between the time the interaction occurred and the current time. This represents the preset time decay coefficient. Indicates the numerical value of the interaction depth level. This represents the depth adjustment factor.

[0008] In the preferred scheme, the specific algorithm steps for extracting structured triplet information in step S2 include: S21. Utilize a large language model to identify the text structure of the patent specification, and locate the background technology paragraph, the specific implementation method paragraph, and the beneficial effect paragraph through paragraph semantic analysis; S22. Perform entity relation extraction, extract the defect description of the prior art from the background technology paragraph and mark it as "technical problem"; The core technical features and parameters for solving the technical problem are extracted from the independent claims and specific embodiments and marked as "technical means"; Extract performance improvement indicators from the beneficial effects paragraphs and experimental data tables and label them as "technical effects"; S23. Perform logical serialization, concatenating "technical issues," "technical means," and "technical effects" according to logical causal relationships to form a serialized generation base, which serves as the context source for the presentation generation.

[0009] In the preferred embodiment, the method for calculating the technological novelty index in step S3 is implemented by introducing a K-Means clustering algorithm optimized by the elbow rule. The steps include: S31. Obtain the set of patents read in the user's history, map each patent to a high-dimensional feature vector, and construct a historical interest sample set; S32. Set the number of clusters The traversal range, for each Perform pre-clustering, calculate the corresponding intra-cluster sum of squared errors (SSE), and plot. The curve is analyzed and its inflection points are identified. The corresponding inflection points are then assigned to the curves. The value is determined as the optimal number of clusters. ; S33. Employ the K-Means++ algorithm strategy to select samples from the historical interest sample set according to probability. We have an initial centroid, ensuring that the distance between the initial centroids is maximized. S34. Calculate the Euclidean distance from each sample to each centroid, assign the sample to the nearest cluster, and recalculate the mean of each cluster as the new centroid. Repeat this process until the centroid positions converge, and output the final result. Centroid vectors of historical interest clusters; S35. Calculate the Euclidean distance between the patent value vector of the patent to be pushed and the centroid vector of the nearest historical interest cluster to the patent, and use it as the novelty distance; S36. The novelty distance is normalized using the Sigmoid function to calculate the technical novelty index. The calculation formula is as follows: ; in, Indicators representing technological novelty This represents the patent value vector of the patent to be pushed. This represents the centroid vector of the historical interest cluster closest to the patent. This represents the Euclidean distance operation. This represents the preset novelty perception threshold. This represents the adjustment coefficient.

[0010] In the preferred embodiment, the method for calculating the competition risk index in step S3 is implemented by introducing the K-Shell decomposition algorithm, and the steps include: S37. Construct a large-scale patent citation network graph that includes competitor nodes of interest to users and a massive number of patent nodes; S38. Perform K-Shell decomposition on the referenced network graph, iteratively removing nodes with a degree less than $ks$, until the degree of all remaining nodes in the network is not less than $ks$. Define the number of levels removed as the kernel value of the node. The network influence coefficient is determined based on the core values ​​of competitor nodes. ; S39. If the applicant of the patent to be pushed is directly in the competitor list, assign the highest risk value; if not, use the breadth-first search algorithm to calculate the minimum path hops between the patent node to be pushed and the competitor node. S310. The competition risk index is calculated by combining the network influence coefficient and the number of path hops. The calculation formula is as follows: ; in, Indicators of competitive risk This indicates the applicant whose patent is to be submitted. This indicates a list of competitor entities. This represents the competitor's network influence coefficient obtained based on K-Shell decomposition. This represents the minimum path hop count. This indicates the maximum number of risk warning jumps.

[0011] In the preferred embodiment, the determination of the target patent to be pushed in step S3 is based on a comprehensive score screening. The calculation sub-steps of this comprehensive score include: S311. Obtain the calculated semantic similarity, technological novelty index, and competition risk index respectively; S312. The three indicators above are linearly weighted and summed using preset or dynamically adjusted weight parameters to obtain a comprehensive score; S313. The formula for calculating the overall score is: ; in, This represents the overall score. The cosine similarity between the user profile vector and the patent value vector is represented. Indicators representing technological novelty Indicators of competitive risk This represents the corresponding weight parameter.

[0012] In the preferred embodiment, the step of invoking the corresponding role attention mask mechanism in step S4 includes: S41. If the user role label is R&D engineer, activate the attention mask of the technical implementation dimension. When generating the large language model, assign high attention weights to the information labeled "technical means" and "specific implementation methods" in the generation basis, and suppress the output of other information. S42. If the user role label is corporate decision-maker, activate the attention mask of the market value dimension. When generating the large language model, assign high attention weights to information marked as "technical effect" and "technical problem" in the generation base, and associate it with the patent family layout information. S43. If the user role is labeled as intellectual property legal counsel, activate the attention mask for the legal protection dimension and assign high attention weights to information marked as "core features of claims" in the generation base during the generation of the large language model.

[0013] In the preferred embodiment, step S5, which involves collecting feedback behavior data and calculating the feedback reward value, includes: S51. Capture user actions on a single briefing and categorize these actions as clicks, favorites, forwards, ignores, and negative feedback; S52. Assign corresponding reward coefficients to different types of operation behaviors, setting click behavior as a positive small value, collection and forwarding behavior as a positive large value, and ignoring and negative feedback behavior as a negative value; S53. The formula for calculating the feedback reward value by summing up the rewards generated by all operations is as follows: ; in, This indicates the feedback reward value. This represents the total number of operations. Indicates the first The occurrence state variable of each operation behavior. This represents the reward coefficient for the corresponding action.

[0014] In the preferred scheme, the dynamic adjustment of weight parameters in step S5 is implemented using the policy gradient algorithm, and the steps include: S54. Construct a policy network, with user profile vectors as input states and weight parameter combinations as action outputs; S55. Based on the calculated feedback reward value, update the parameters of the policy network using the gradient ascent method to increase the probability of high reward weight combinations being selected; S56. The formula for updating the weight parameters is: ; in, Represents the set of control weight parameters in the policy network. Network parameters, Indicates the learning rate. Indicates the current user profile status. Select a specific weight combination The probability, This indicates the feedback reward value.

[0015] A patent technology briefing intelligent push system based on user profiles and large models includes: a profile building module, used to acquire user data, calculate the weight of technical topics using time decay and interaction depth weighting algorithms, and build a user profile vector containing user role tags and technical interest distribution; The semantic parsing module is used to semantically decompose patent documents using a large language model that has been fine-tuned with domain knowledge, extract the structured triple information of "technical problem-technical means-technical effect" that represents the essence of the patent and the core features of the claims, and map them into vectors; The intelligent matching module is used to calculate semantic similarity, combine the technical novelty index calculated by K-Means clustering algorithm based on elbow rule optimization, and the competition risk index calculated by K-Shell decomposition algorithm, and use the multi-factor linear weighted summation formula to calculate the comprehensive score and screen target patents. The briefing generation module is used to call the corresponding role attention masking mechanism based on the user's role label, and use a large language model to reorganize the structured triple information to generate personalized technical briefings. The feedback optimization module is used to collect user feedback data, quantify user feedback using the feedback reward value accumulation formula, and dynamically update the weight parameters in the intelligent matching module based on the reward value using the policy gradient algorithm.

[0016] This invention provides a method and system for intelligently pushing patent technology briefings based on user profiles and large-scale models. By introducing a dynamic profile construction mechanism based on time decay and interaction depth weighting, this invention can adjust the weight of technical topics in real time according to the time proximity and depth of user behavior, thereby more accurately capturing the user's current true technical interests and avoiding invalid pushes due to outdated data. By utilizing a large language model to extract structured triples of "technical problem-technical means-technical effect" and combining them with a role attention masking mechanism, this invention can reorganize the same patent into personalized briefings with different focuses for different roles such as R&D engineers, corporate decision-makers, and intellectual property lawyers. For example, it can show specific technical parameters to R&D personnel and market value to decision-makers, greatly reducing the information cognitive load and reading threshold for different roles.

[0017] In the intelligent matching stage, this invention innovatively introduces a K-Means clustering algorithm optimized based on the elbow rule to calculate the novelty index of technology, effectively breaking the "information cocoon" effect of traditional recommendation systems and helping users discover inspiring cross-industry technologies. Simultaneously, it utilizes the K-Shell decomposition algorithm to deeply analyze the patent citation network, accurately quantifying competitors' network influence and potential infringement risks, providing enterprises with strategically significant competitive intelligence warnings. Furthermore, the system constructs a closed-loop feedback mechanism through a policy gradient reinforcement learning algorithm, dynamically and adaptively adjusting recommendation weights based on users' actual feedback behavior. This ensures that the push strategy continuously evolves with changes in user needs, achieving a leap from simple information distribution to high-value intelligent intelligence services. Attached Figure Description

[0018] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a flowchart of the intelligent push method for the patented technology briefing of this invention; Figure 2 This is a software interface diagram of the intelligent push method for the technology briefing of this invention. Detailed Implementation

[0019] Example 1 like Figure 1-2 As shown, a method for intelligently pushing patent technology briefings based on user profiles and large models includes: S1. Obtain explicit user identity information and implicit interaction behavior data, and dynamically weight the user's technical interest characteristics based on the time proximity and interaction depth of the user behavior to obtain a user profile vector containing user role tags and technical interest distribution. S2. Using a large language model fine-tuned with domain knowledge, semantically decompose the patent documents to be pushed, extract the structured triple information of "technical problem, technical means, and technical effect" that can represent the essence of the patent technology, as well as the core features of the claims, and map the extracted information into a patent value vector; S3. Calculate the semantic similarity between the user profile vector and the patent value vector, and introduce the technical novelty index and the competition risk index to weight and correct the semantic similarity. Based on the weighted comprehensive score, determine the target patent to be pushed. S4. Using the structured triplet information of the target patent as the generation basis, based on the user role label in the user profile vector, call the corresponding role attention mask mechanism, use the large language model to differentiate and rewrite the generation basis, generate a personalized technical briefing adapted to the user role and push it. S5. Collect user feedback data on the push notifications and dynamically adjust the weight parameters of the weighted correction in S3 using reinforcement learning algorithms.

[0020] This method first aims to construct a highly dynamic and accurate user profile. In the steps of acquiring explicit user identity information and implicit interaction behavior data, the system does not rely solely on static registration information but updates the profile by capturing the user's interaction behavior flow in real time. Specifically, the system obtains the timestamp of the user's behavior and calculates the time difference between that timestamp and the current moment. It then uses an exponential decay function or linear decay logic to calculate the time decay coefficient corresponding to this time difference, thus assigning greater weight to behaviors closer to the current moment, ensuring that the profile reflects the user's latest technological interests. Simultaneously, the system sets different depth levels based on the type of interaction behavior, assigning high weight to deep interactions such as downloading full texts, collecting patents, or frequently forwarding content, and low weight to shallow interactions such as merely browsing titles, simple searches, or brief pauses. By multiplying the base weight by the time decay coefficient, the system can accurately calculate the effective contribution value of a single behavior and accumulate the effective contribution values ​​of all single behaviors under the same technical topic, thereby generating a multi-dimensional user profile vector that includes user role tags and the distribution of technical interests.

[0021] In the patent document processing stage, this invention utilizes a domain-knowledge-tuned large language model to perform deep semantic decomposition of the patent documents to be pushed. This process aims to address the information loss problem in traditional abstract generation methods. The large language model is trained to recognize the text structure of the patent specification, accurately locating key paragraphs such as background technology, specific implementation methods, and beneficial effects through paragraph semantic analysis. Based on this, the model performs entity relation extraction, extracting descriptions of prior art defects as technical problems from the background technology paragraphs, extracting core technical features and parameters for solving these technical problems as technical means from the independent claims and specific implementation methods, and extracting performance improvement indicators as technical effects from the beneficial effects paragraphs and experimental data tables. These three elements together constitute structured triple information representing the essence of the patent technology. Subsequently, the system maps the extracted triple information and core features of the claims into a high-dimensional patent value vector, providing a data foundation for subsequent mathematical operations.

[0022] In the intelligent matching and ranking stage, the system calculates the semantic similarity between the user profile vector and the patent value vector, using this as the basic recommendation score. To break down information silos and provide strategic early warnings, the system further introduces a technology novelty index and a competition risk index to weight and correct the semantic similarity. The technology novelty index is calculated by obtaining the user's historical patent reading set and using a clustering algorithm to divide it into historical interest clusters. The distance between the patent to be recommended and the centroid of the most recent historical interest cluster is calculated; the greater the distance, the higher the novelty. The competition risk index is determined by constructing a patent citation network graph and analyzing the number of citation path hops or the coreness of network nodes between the patent applicant to be recommended and the user's competitors. The system determines the target patents to be recommended based on the weighted comprehensive score, ensuring that the recommendation results not only match user interests but also have forward-looking and strategic value.

[0023] In the key steps of briefing generation, this invention employs a role-based attention mask mechanism to generate personalized content. The system uses the structured triplet information of the target patent as the generation basis and calls the corresponding attention mask based on the user role label in the user profile vector. For example, when the user role is a research and development engineer, the system activates the attention mask for the technical implementation dimension, forcing the large language model to allocate higher attention weights to the technical means and parameter details in the specific implementation during generation; when the user role is a corporate decision-maker, the system activates the attention mask for the market value dimension, focusing on the technical effects, the pain points solved, and the distribution of patent families; when the user role is an intellectual property lawyer, the system activates the attention mask for the legal protection dimension, focusing on the scope of protection of the claims and the background of the technical problem. Through this differentiated recombination and rewriting, the system can generate personalized technical briefings that are fully adapted to the user role and push them out.

[0024] Finally, to ensure the system's continuous evolution, it collects user feedback data on push notifications. This feedback data is quantified into reward values; for example, clicking corresponds to a weak positive reward, saving corresponds to a strong positive reward, and ignoring corresponds to a negative reward. Using reinforcement learning algorithms such as policy gradient, the system treats user profiles as states and weight parameter adjustments as actions. It updates the policy network through backpropagation based on feedback reward values, thereby dynamically adjusting the weight parameters of semantic similarity, novelty, and competition risk in the overall score. This allows the push strategy to adaptively optimize as user needs change.

[0025] The technical solution provided by this invention has significant beneficial effects. First, by introducing a dynamic profile construction mechanism that weights time decay and interaction depth, this invention can accurately capture users' changing interests over time and distinguish the intensity of real needs behind different behaviors, significantly improving the timeliness and accuracy of user profiles. Second, by utilizing a large language model to extract structured triples and combining them with a role attention masking mechanism, this invention solves the problem of traditional patent abstracts being monotonous and difficult to understand. It can provide customized briefings with clear focus for different roles such as R&D, decision-making, and legal affairs, greatly reducing users' information acquisition costs and cognitive load. Third, by introducing technical novelty indicators and competitive risk indicators in the matching process, this invention can not only help users discover inspiring cross-border technologies and break information silos, but also effectively identify potential competitors and infringement risks, enhancing the strategic reference value of patent intelligence. Finally, the closed-loop feedback mechanism based on reinforcement learning ensures that system parameters can be automatically iteratively optimized according to actual user feedback, achieving continuous improvement in push accuracy and long-term improvement in user experience.

[0026] In the preferred scheme, the dynamic weighting of user technical interest features in step S1 is implemented based on a time decay and interaction depth weighting algorithm, and the steps are as follows: S11. Preset a basic weight system, setting download and bookmark behaviors as high weight levels and browsing and searching behaviors as low weight levels to determine the basic weight of a single interaction behavior; S12. Obtain the timestamp of the interaction, calculate the time difference between the timestamp and the current time, and use the exponential decay logic to calculate the time decay coefficient corresponding to the time difference, so that the closer the behavior is to the current time, the larger the retention coefficient. S13. Determine the depth adjustment factor based on the interaction depth level, correct the basic weight, and multiply the corrected basic weight by the time decay coefficient to obtain the effective contribution value of a single behavior. S14. Based on the above steps, calculate the final weight of the specific technical topic, using the following formula: ; in, This indicates the final weight of the technical topic. The base weight representing a single interaction behavior. This represents the time difference between the time the interaction occurred and the current time. This represents the preset time decay coefficient. Indicates the numerical value of the interaction depth level. This represents the depth adjustment factor.

[0027] In this embodiment of the invention, a specific implementation method is provided for the dynamic weighting process of user technical interest characteristics described in step S1 and its sub-steps S11 to S14. This method first establishes a multi-dimensional weight evaluation system to quantify the behavioral value of users interacting with the patent system. In the stage of pre-setting the basic weight system, the system classifies users according to the strength of their subjective intentions reflected in their behavior. Behaviors demonstrating a strong desire to acquire information, such as downloading the full text of a patent and collecting patent documents, are defined as high-weighted and assigned larger base values, while exploratory behaviors such as regular page browsing and keyword searching are defined as low-weighted and assigned smaller base values, thereby determining the basic weight of a single interaction. Based on this, the system introduces a time dimension to simulate the natural decay law of user interest. By obtaining the precise timestamp of each interaction, the system calculates the time difference between this timestamp and the current system moment. To reflect the high correlation between recent behavior and current interest, the system uses exponential decay logic to process this time difference, calculating a time decay coefficient. This coefficient decreases exponentially as the time difference increases, ensuring that behaviors closer to the current moment have a larger retention coefficient in the final calculation, while the influence of distant historical behaviors is rapidly weakened.

[0028] Furthermore, the system integrates interaction depth with time factors in its calculations. Based on a pre-determined interaction depth level, the system matches a corresponding depth adjustment factor, using this factor to enhance or correct the base weights, reflecting the core role of deep interaction behavior in interest determination. Subsequently, the corrected weight value is multiplied by the calculated time decay coefficient to obtain the effective contribution value of this single behavior to the user's interest profile. To comprehensively evaluate the user's interest concentration in a specific technical field, the system aggregates the effective contribution values ​​of all interactions generated by the user under that specific technical topic, and finally calculates the final weight of that technical topic through a mathematical model. This calculation process strictly follows the formula. .

[0029] In this mathematical model, each parameter has a clear technical meaning. This represents the final weight value of the technology topic obtained after calculation, and is used to directly measure the degree of user attention to this technology field. This represents the base weight value for a single interaction behavior, determined based on the behavior type. (Mathematical notation) It represents the base of the natural logarithm. This represents the preset time decay coefficient, which is used to control the rate of forgetting. The larger the value, the more sensitive it is to recent behavior. This represents the time difference between when the interaction occurred and the current moment. This represents the depth adjustment factor, used to adjust the gain of interaction depth on the total weight. This represents the specific numerical value corresponding to the level of interaction depth. (Symbol) This represents the summation of all historical interaction records under this technical topic.

[0030] The above technical solution yields significant benefits. By introducing a dual weighted mechanism of time decay and interaction depth, this invention effectively solves the technical problem that static tags in traditional user profile construction cannot reflect the time-varying nature of interests. First, using exponential decay logic, the system can automatically filter outdated interest tags, ensuring that the pushed content keeps up with the user's latest research and development directions or hot topics, avoiding interference from outdated information. Second, by quantitatively distinguishing interaction depth, the system can accurately identify the intensity of the user's true needs, preventing profile distortion caused by high-frequency but low-value misoperations.

[0031] In the preferred scheme, the specific algorithm steps for extracting structured triplet information in step S2 include: S21. Utilize a large language model to identify the text structure of the patent specification, and locate the background technology paragraph, the specific implementation method paragraph, and the beneficial effect paragraph through paragraph semantic analysis; S22. Perform entity relation extraction, extract the defect description of the prior art from the background technology paragraph and mark it as "technical problem"; The core technical features and parameters for solving the technical problem are extracted from the independent claims and specific embodiments and marked as "technical means"; Extract performance improvement indicators from the beneficial effects paragraphs and experimental data tables and label them as "technical effects"; S23. Perform logical serialization, concatenating "technical issues," "technical means," and "technical effects" according to logical causal relationships to form a serialized generation base, which serves as the context source for the presentation generation.

[0032] This invention provides a detailed implementation method and technical effect description for the deep semantic decomposition and structured information extraction process of patent documents described in step S2 and its sub-steps S21 to S23. In the step of recognizing the text structure of the patent specification using a large language model, the system does not simply rely on keyword matching, but utilizes the contextual understanding capability of the large language model to perform paragraph-level semantic analysis on the entire patent text. By identifying semantic boundaries in the specification, the system can accurately locate background technology paragraphs describing the defects of the prior art, specific implementation paragraphs detailing the technical solutions, and beneficial effect paragraphs illustrating the technical advantages. Based on this, the system performs a fine-grained entity relationship extraction task. For the background technology paragraphs, the model focuses on identifying descriptive statements about the shortcomings, pain points, or unresolved needs of the prior art, marking them as technical problems. For the independent claims and specific implementation paragraphs, the model extracts the core structural features, process step parameters, and key algorithm logic used to solve the above-mentioned technical problems through syntactic analysis, marking them as technical means. For the beneficial effect paragraphs and experimental data tables in the accompanying drawings, the model extracts specific performance improvement values, efficiency optimization ratios, or functional enhancement descriptions, marking them as technical effects.

[0033] After extracting the aforementioned information elements, the system performs a logical serialization operation. This process is not a simple text concatenation, but rather a logical causal chain of "problems lead to needs, methods solve problems, and effects verify methods," systematically reorganizing technical problems, technical methods, and technical effects. The system defines this logically structured data as the generation basis for serialization and uses it as the direct contextual source for the subsequent generation of personalized briefings by the large language model. This ensures that the generation model can rewrite based on accurate and logically rigorous factual data, rather than generating uncontrollable summaries based on vague full text.

[0034] First, by clearly defining the extraction of specific technical elements from specific paragraphs such as "Background Technology," "Specific Implementation Methods," and "Beneficial Effects," this invention transforms unstructured long-text patents into clearly structured triple information through technical means. This fundamentally solves the "illusion" or focus shift problems that easily occur when large language models process long texts, ensuring the accuracy and traceability of extracted information. Second, the logically sequential processing method restores the thought process of invention, enabling the subsequently generated presentations to not only include "what" but also clearly explain "why" and "how," significantly improving the readability and logic of the presentations.

[0035] In the preferred embodiment, the method for calculating the technological novelty index in step S3 is implemented by introducing a K-Means clustering algorithm optimized by the elbow rule. The steps include: S31. Obtain the set of patents read in the user's history, map each patent to a high-dimensional feature vector, and construct a historical interest sample set; S32. Set the number of clusters The traversal range, for each Perform pre-clustering, calculate the corresponding intra-cluster sum of squared errors (SSE), and plot. The curve is analyzed and its inflection points are identified. The corresponding inflection points are then assigned to the curves. The value is determined as the optimal number of clusters. ; S33. Employ the K-Means++ algorithm strategy to select samples from the historical interest sample set according to probability. We have an initial centroid, ensuring that the distance between the initial centroids is maximized. S34. Calculate the Euclidean distance from each sample to each centroid, assign the sample to the nearest cluster, and recalculate the mean of each cluster as the new centroid. Repeat this process until the centroid positions converge, and output the final result. Centroid vectors of historical interest clusters; S35. Calculate the Euclidean distance between the patent value vector of the patent to be pushed and the centroid vector of the nearest historical interest cluster to the patent, and use it as the novelty distance; S36. The novelty distance is normalized using the Sigmoid function to calculate the technical novelty index. The calculation formula is as follows: ; in, Indicators representing technological novelty This represents the patent value vector of the patent to be pushed. This represents the centroid vector of the historical interest cluster closest to the patent. This represents the Euclidean distance operation. This represents the preset novelty perception threshold. This represents the adjustment coefficient.

[0036] This embodiment of the invention addresses the calculation process of the technical novelty index described in step S3 and its sub-steps S31 to S36. This method does not rely on simple keyword matching or fixed classification labels to determine novelty; instead, it employs dynamic clustering analysis based on unsupervised learning. In the historical data processing stage, the system first obtains the set of patents the user has historically read. Using a pre-trained model, each unstructured patent text is mapped into a high-dimensional feature vector that can be processed by a computer, thereby constructing a historical interest sample set that can represent the user's past technical preferences. To address the problem of determining the number of clusters in traditional clustering algorithms, the system introduces the elbow rule for dynamic optimization. The system changes the number of clusters within a preset traversal range. The value is then used to perform pre-clustering and calculate the corresponding sum of squared errors within clusters (SSE). Through analysis... The system identifies the inflection point where the rate of decrease in the curve changes abruptly from the curve's decreasing rate, and determines the value corresponding to this inflection point as the optimal number of clusters. This step ensures that the system can automatically divide users into the most appropriate number of interest sections based on the breadth of their interests, rather than using rigid, fixed categories.

[0037] During the clustering execution phase, the system employs the K-Means++ algorithm to probabilistically select initial centroids from the sample set. This strategy ensures that the distance between initial centroids is maximized, effectively avoiding the algorithm getting trapped in local optima, thereby improving the stability and accuracy of clustering. Subsequently, the system iteratively calculates the Euclidean distance from each sample to each centroid, assigns the sample to the nearest cluster, and updates the cluster's average value until the centroid positions converge. Finally, it outputs the centroid vectors of historical interest clusters representing the user's core interest domain. Based on this, the system calculates the Euclidean distance between the patent value vector of the patent to be pushed and the centroid vector of the nearest historical interest cluster to that patent. This distance intuitively quantifies the degree of difference between the patent to be pushed and the user's previously familiar technical field, i.e., the novelty distance. To transform this physical distance into a standardized metric that can be used for weighted calculation, the system uses the Sigmoid function for nonlinear normalization processing. The specific calculation model is as follows: .

[0038] In this mathematical model, each parameter has a clear physical meaning. This represents the final calculated technical novelty index. The higher the value, the more novel and inspiring the patent is to current users. This represents the value vector of the patent to be pushed after vectorization. This represents the centroid vector of the cluster closest to the patent to be pushed, across all of the user's historical interest clusters. (Symbol) This represents the Euclidean distance operation, used to measure absolute distance in a vector space. This represents a preset novelty perception threshold. When the distance exceeds this threshold, the novelty index will increase significantly. This represents the adjustment coefficient, used to control the slope of the Sigmoid function curve, that is, to control the sensitivity of the index to changes in distance.

[0039] First, by introducing K-Means clustering optimized by the elbow rule, this invention can adaptively identify the multimodal interest distribution of users, avoiding the subjectivity of manually setting the number of clusters and making the definition of users' "familiar domains" more scientific and objective. Second, by calculating the distance between the patent to be recommended and the nearest centroid of interest and performing Sigmoid normalization, the system can finely control the novelty of the recommendations.

[0040] In the preferred embodiment, the method for calculating the competition risk index in step S3 is implemented by introducing the K-Shell decomposition algorithm, and the steps include: S37. Construct a large-scale patent citation network graph that includes competitor nodes of interest to users and a massive number of patent nodes; S38. Perform K-Shell decomposition on the referenced network graph, iteratively removing nodes with a degree less than $ks$, until the degree of all remaining nodes in the network is not less than $ks$. Define the number of levels removed as the kernel value of the node. The network influence coefficient is determined based on the core values ​​of competitor nodes. ; S39. If the applicant of the patent to be pushed is directly in the competitor list, assign the highest risk value; if not, use the breadth-first search algorithm to calculate the minimum path hops between the patent node to be pushed and the competitor node. S310. The competition risk index is calculated by combining the network influence coefficient and the number of path hops. The calculation formula is as follows: ; in, Indicators of competitive risk This indicates the applicant whose patent is to be submitted. This indicates a list of competitor entities. This represents the competitor's network influence coefficient obtained based on K-Shell decomposition. This represents the minimum path hop count. This indicates the maximum number of risk warning jumps.

[0041] This invention provides a detailed implementation method for calculating the competition risk index in step S3 and its sub-steps S37 to S310. This method is not limited to simple competitor list matching, but rather deeply mines the implicit risks in patent citation relationships based on complex network theory. First, the system constructs a large-scale patent citation network graph. This graph not only includes competitor nodes that the user is explicitly concerned with, but also covers a massive number of related patent nodes. The connections between nodes are constituted by citation and cited relationships determined by patent citation data. Based on this, the system executes the K-Shell decomposition algorithm to evaluate the core influence of each node in the network. The algorithm proceeds iteratively, first stripping edge nodes with a degree less than a specific threshold $ks$ and assigning these nodes to their corresponding levels. As edge nodes are removed, the degree of the remaining nodes changes. The system repeats this process until the degree of all remaining nodes in the network is not less than the current threshold $ks$. At this point, the number of levels stripped is defined as the core value of that node. The system determines the network influence coefficient of competitor nodes based on their final core values ​​through normalization or a mapping function. This coefficient objectively reflects whether a competitor occupies a core or peripheral position in the technology network.

[0042] After completing the node influence assessment, the system enters the specific risk determination stage. The system first determines whether the applicant of the patent to be pushed directly exists in a pre-defined list of competitor entities. If so, it indicates that the patent originates directly from a competitor, possessing the highest competitive intelligence value or infringement risk, and is therefore assigned the highest risk value. If the applicant is not on the list, the system uses a breadth-first search algorithm to calculate the shortest path, i.e., the minimum path hop count, between the patent node to be pushed and known competitor nodes in the citation network graph. This step aims to discover potentially risky patents that, while not directly applied for by competitors, are highly related to competitor technology through citation relationships. Finally, the system combines the network influence coefficient and the path hop count using a formula... Calculate the competition risk index. In this formula, This represents the final competitive risk indicator value; Information about the applicant for the patent to be pushed to the system; This represents the collection of competitor entities stored in the system. This represents the competitor's technological influence coefficient calculated based on the K-Shell decomposition algorithm, used to weight path risk; This represents the minimum number of citation path jumps between the patent to be pushed and its competitors; the smaller the number of hops, the more direct the risk transmission. This represents the maximum number of risk warning jumps preset by the system. Associations exceeding this number of jumps are considered to have negligible risk.

[0043] First, by introducing the K-Shell decomposition algorithm, this invention overcomes the limitations of traditional methods that rely solely on citation counts to assess influence. It accurately identifies high-risk nodes located at the core of the technology network, distinguishing the threat levels of core competitors from peripheral competitors, thus making risk assessment more comprehensive and objective. Second, by using a breadth-first search algorithm to calculate hop counts in citation paths, the system can keenly detect hidden technological risks, such as patents applied for by competitors through subsidiaries or partners, or third-party patents with highly overlapping technological paths with competitors, thereby providing users with penetrating intelligence alerts.

[0044] In the preferred embodiment, the determination of the target patent to be pushed in step S3 is based on a comprehensive score screening. The calculation sub-steps of this comprehensive score include: S311. Obtain the calculated semantic similarity, technological novelty index, and competition risk index respectively; S312. The three indicators above are linearly weighted and summed using preset or dynamically adjusted weight parameters to obtain a comprehensive score; S313. The formula for calculating the overall score is: ; in, This represents the overall score. The cosine similarity between the user profile vector and the patent value vector is represented. Indicators representing technological novelty Indicators of competitive risk This represents the corresponding weight parameter.

[0045] This invention provides specific implementation methods and technical effects for the comprehensive score calculation and target patent screening process described in step S3 and its sub-steps S311 to S313. In the core stage of multi-factor intelligent matching and ranking, the system first performs a data aggregation operation, retrieving three core indicator data calculated in previous steps from memory or a temporary database. These three indicators are: semantic similarity (representing the semantic overlap between user interests and patent content), technical novelty (representing the degree of difference between the patent's technical solution and the user's historical knowledge base), and competitive risk (representing the degree of potential commercial competitive threat). After obtaining the above basic data, the system uses a linear weighted summation model to fuse multi-dimensional features, adjusting the importance of indicators in different dimensions by introducing weight parameters to calculate the final recommendation score for the patent to be pushed. The specific mathematical model is expressed as follows: .

[0046] In this mathematical calculation model, each variable has a clear and independent physical meaning and technical orientation. This represents the final calculated comprehensive score, which is the sole basis for determining whether a patent enters the final push queue and its order within the queue. The cosine similarity value represents the difference between the user profile vector and the patent value vector, and is used to ensure the basic relevance of the recommended content. This represents a normalized value for the technical novelty index, used to measure the inspiration of a technical solution. This represents a competitive risk indicator value derived from network analysis, used to measure the strategic value of intelligence. , and These represent the weight parameters corresponding to the three indicators mentioned above. The sum of these three weight parameters is usually set to 1, and their specific values ​​can be either fixed values ​​preset by the system or variables that are dynamically adjusted based on user feedback through reinforcement learning algorithms in subsequent steps.

[0047] First, by constructing a multi-factor comprehensive evaluation system that includes semantic similarity, technological novelty, and competitive risk, this invention completely solves the "information cocoon" problem caused by traditional recommendation systems relying solely on a single similarity index. This computational model not only considers what users "want to see," but also takes into account the novel technologies users "need to see" and the competitive risks they "must be wary of," making the recommended results relevant, exploratory, and strategic. Second, by using a linear weighting formula to fuse multi-dimensional heterogeneous data into a single sortable value, it greatly simplifies the decision-making logic and improves the system's computational efficiency and ranking stability. Finally, clearly defined weight parameters... , , It provides an interface for the system's flexibility, enabling the system to focus on different recommendation dimensions based on different user roles or usage scenarios. For example, it can increase the novelty weight for R&D personnel and the risk weight for legal personnel, thereby achieving true intelligent and personalized push. These detailed parameter definitions and logical descriptions provide sufficient support for the features of comprehensive score calculation in the claims.

[0048] In the preferred embodiment, the step of invoking the corresponding role attention mask mechanism in step S4 includes: S41. If the user role label is R&D engineer, activate the attention mask of the technical implementation dimension. When generating the large language model, assign high attention weights to the information labeled "technical means" and "specific implementation methods" in the generation basis, and suppress the output of other information. S42. If the user role label is corporate decision-maker, activate the attention mask of the market value dimension. When generating the large language model, assign high attention weights to information marked as "technical effect" and "technical problem" in the generation base, and associate it with the patent family layout information. S43. If the user role is labeled as intellectual property legal counsel, activate the attention mask for the legal protection dimension and assign high attention weights to information marked as "core features of claims" in the generation base during the generation of the large language model.

[0049] This invention provides a detailed implementation method and technical principle explanation for the role attention masking mechanism call and personalized briefing generation process described in step S4 and its sub-steps S41 to S43. The core of this method lies in utilizing the attention mechanism of a large language model to dynamically adjust the model's attention to different parts of the input sequence during the decoding and generation process, thereby achieving differentiated output of the same source document. In specific implementation, the system pre-constructs an attention mask matrix or prompt word strategy corresponding to the user's role label. When the system identifies the current user's role label as a research engineer, it automatically activates the attention mask for the technical implementation dimension. This masking mechanism significantly increases the weight parameters of text fragments corresponding to technical means and specific implementation methods when the model processes the input sequence, making them dominant in self-attention calculation, while reducing the weight of non-technical information such as background introductions or legal status, thereby suppressing the output of these interfering information. Based on this, the large language model generates in-depth technical briefings focusing on technical parameters, process flow diagrams, and core algorithm logic, helping R&D personnel quickly understand the technical essence.

[0050] For corporate decision-makers, the system activates an attention mask for the market value dimension. This mechanism guides the large language model to allocate high attention weights to the parts of the generated base labeled as technical effects and technical problems, focusing on uncovering the pain points that patents aim to solve and the performance improvements or cost advantages brought about by technological advancements. Furthermore, this step involves the correlation of external data; the system introduces retrieved patent family layout information into the context window and assigns it high weight. This ensures that the model-generated briefings not only include technical value analysis but also incorporate an assessment of the global market competitive landscape based on patent family distribution, providing decision-makers with a macro-level perspective for their business strategy.

[0051] When the user's role is identified as intellectual property legal counsel, the system switches to an attention mask focused on legal protection. At this point, the large language model's attention is forcibly locked onto the core features of the claims and the description of the scope of protection of the independent claims. The system uses a masking mechanism to block out a large amount of detailed description of embodiments, prompting the model to focus on analyzing the boundaries of the claims, the composition of essential technical features, and potential infringement risks, thereby generating a rigorously worded risk assessment report that emphasizes the definition of the scope of legal protection.

[0052] First, this invention innovatively applies attention masking technology from large language models to the personalized reorganization of patent intelligence, completely changing the traditional one-size-fits-all approach to patent services. By controlling the information flow at the underlying algorithmic level, the system ensures that the generated briefing content accurately matches the cognitive habits and work needs of different professional roles, greatly reducing the time cost for users to sift through lengthy patent documents for key information. Second, the differentiated masking strategy for different roles not only improves the efficiency of information transmission but also ensures the professional depth of the information, such as providing reproducible technical details for R&D personnel, business insights for decision-makers, and risk control basis for legal personnel.

[0053] In the preferred embodiment, step S5, which involves collecting feedback behavior data and calculating the feedback reward value, includes: S51. Capture user actions on a single briefing and categorize these actions as clicks, favorites, forwards, ignores, and negative feedback; S52. Assign corresponding reward coefficients to different types of operation behaviors, setting click behavior as a positive small value, collection and forwarding behavior as a positive large value, and ignoring and negative feedback behavior as a negative value; S53. The formula for calculating the feedback reward value by summing up the rewards generated by all operations is as follows: ; in, This indicates the feedback reward value. This represents the total number of operations. Indicates the first The occurrence state variable of each operation behavior. This represents the reward coefficient for the corresponding action.

[0054] This invention provides a detailed implementation description of the feedback data collection and reward value calculation process described in step S5 and its sub-steps S51 to S53. This step constructs a quantitative evaluation system based on explicit user behavior to provide accurate reward signals for reinforcement learning algorithms, thereby driving the self-evolution of recommendation strategies. In the operation behavior capture stage, the system monitors the user's interaction with the push notification in real time through front-end tracking or log analysis technology. The system abstracts and categorizes complex interactive behaviors into five core operation types: click to read, save, forward and share, ignore directly, and proactive negative feedback. These behavior types objectively reflect the user's satisfaction and acceptance of the notification content. To transform discrete behavioral data into scalar values ​​that the reinforcement learning model can understand, the system establishes a differentiated reward coefficient allocation mechanism. The system considers clicks to represent a user's initial interest, therefore setting its corresponding reward coefficient to a small positive value. Favorites and shares represent a user's high level of recognition and in-depth need for content value, thus setting their corresponding reward coefficient to a large positive value to encourage the model to recommend more such high-value content. Ignoring or providing negative feedback is considered invalid recommendations or interference, therefore its reward coefficient is set to a negative value as a penalty for the recommendation strategy. Based on this, the system uses an additive summation algorithm to calculate the total feedback reward value within a specific recommendation batch or time window, a calculation process strictly adhering to a mathematical model. implement.

[0055] In this mathematical model, each parameter has a clear definition and calculation logic. This represents the final calculated feedback reward value, which is the core basis for the reinforcement learning algorithm to update the policy network parameters. The higher the value, the more the current recommendation strategy meets the user's needs. This represents the total number of user actions during the statistical period, ensuring that reward calculations cover all interaction records. Representing the The occurrence status variable of an operation behavior, usually takes the value of 1 or 0, and is used to indicate whether a specific behavior has occurred. Represents the corresponding number The pre-set reward coefficient for each action directly determines the direction and extent of that action's contribution to the total reward, whether it is positive or negative and its absolute value.

[0056] First, by establishing a refined behavior classification and differentiated reward mechanism, this invention overcomes the limitations of traditional recommendation systems that solely optimize click-through rate (CTR). By assigning higher weights to favorites and shares, the system guides the model from simply pursuing "eye-catching" content to pursuing "content value," effectively preventing the proliferation of clickbait content and enhancing the professionalism and user stickiness of the push service. Second, the introduction of a negative penalty mechanism enables the model to quickly identify and filter content types that users are not interested in or dislike, accelerating the algorithm's convergence speed. Finally, a clear mathematical formula transforms abstract user satisfaction into a calculable and optimizable mathematical objective function, achieving a closed-loop connection between user feedback and model updates. In the preferred scheme, the dynamic adjustment of weight parameters in step S5 is implemented using the policy gradient algorithm, and the steps include: S54. Construct a policy network, with user profile vectors as input states and weight parameter combinations as action outputs; S55. Based on the calculated feedback reward value, update the parameters of the policy network using the gradient ascent method to increase the probability of high reward weight combinations being selected; S56. The formula for updating the weight parameters is: ; in, Represents the set of control weight parameters in the policy network. Network parameters, Indicates the learning rate. Indicates the current user profile status. Select a specific weight combination The probability, This indicates the feedback reward value.

[0057] This invention provides specific implementation methods and technical principle explanations for the dynamic adjustment of weight parameters described in step S5 and its sub-steps S54 to S56. This step is the core link in realizing the adaptive evolution of the intelligent recommendation system, aiming to solve the problem that fixed weight parameters in traditional recommendation systems cannot adapt to user interest drift and diversified needs. In the stage of constructing the policy network, the system establishes a decision model based on a neural network. This policy network takes the current user profile vector as the input state. This vector contains the user's time-decay interest features and role label information. The network's output layer corresponds to the action space. That is, the set of weight parameters used for multi-factor intelligent matching and ranking. The probability distribution or specific values ​​of semantic similarity, novelty index, and competition risk index are no longer manually set by the system. Instead, the policy network determines the optimal ratio of these three factors in real time based on the current user's specific state.

[0058] After completing a push notification and receiving a user feedback reward, the system updates the network parameters using a policy gradient algorithm. This process follows the optimization logic of gradient ascent, the core idea of ​​which is to adjust the parameters of the policy function based on the reward signal. This increases the probability of weight combinations that can obtain high feedback rewards being selected when encountering similar user profile states in the future, while decreasing the probability of weight combinations that result in negative feedback being selected. The specific parameter update iterations strictly follow mathematical formulas. .

[0059] In this formula, each variable has a clear mathematical definition and physical meaning. This represents all trainable network parameters in the policy network, which determine the network's mapping from input states to output actions. and These represent the network parameters before and after the update, respectively. This represents the learning rate, used to control the step size of parameter updates in order to balance convergence speed and training stability. The representative policy function is the function defined by the current network parameters. Under the influence of specific user profile states When selecting a specific weight combination The probability value. (Symbol) Represents the parameter Find the gradient operator for partial derivatives. This represents the feedback reward value calculated in the preceding steps. This formula indicates that the direction of parameter updates is along the direction of the logarithmic policy gradient, and is determined by the reward value. The magnitude and sign are weighted and the direction is controlled.

[0060] First, this invention transforms the problem of adjusting recommendation weights into a sequence decision-making problem in reinforcement learning, achieving personalized and dynamically evolving push strategies. The system can automatically learn the preference patterns of different user roles at different stages. For example, the system may automatically learn that newly registered R&D engineers should be assigned higher weights for semantic similarity to ensure accuracy, while highly active corporate decision-makers should be assigned higher weights for novelty and risk indicators to provide strategic value. Second, through the closed-loop feedback of the policy gradient algorithm, the system possesses strong robustness and adaptability. It can continuously self-correct and optimize as user feedback data accumulates without human intervention, thus consistently maintaining high-quality recommendation results and high user satisfaction.

[0061] Example 2 Further explanation in conjunction with Example 1, such as Figure 1-2 As shown, a patent technology briefing intelligent push system based on user profiles and large models includes: a profile construction module, used to acquire user data, calculate the weight of technical topics using time decay and interaction depth weighting algorithms, and construct a user profile vector containing user role tags and technical interest distribution; The semantic parsing module is used to semantically decompose patent documents using a large language model that has been fine-tuned with domain knowledge, extract the structured triple information of "technical problem-technical means-technical effect" that represents the essence of the patent and the core features of the claims, and map them into vectors; The intelligent matching module is used to calculate semantic similarity, combine the technical novelty index calculated by K-Means clustering algorithm based on elbow rule optimization, and the competition risk index calculated by K-Shell decomposition algorithm, and use the multi-factor linear weighted summation formula to calculate the comprehensive score and screen target patents. The briefing generation module is used to call the corresponding role attention masking mechanism based on the user's role label, and use a large language model to reorganize the structured triple information to generate personalized technical briefings. The feedback optimization module is used to collect user feedback data, quantify user feedback using the feedback reward value accumulation formula, and dynamically update the weight parameters in the intelligent matching module based on the reward value using the policy gradient algorithm.

[0062] First, the user profile building module is primarily responsible for capturing the shift in user interests over time. This module goes beyond simply recording user history; it executes a time-decay and interaction depth-weighted algorithm. In practice, this module reads user activity logs such as search and downloads, calculates an exponential decay coefficient based on the time of the action, and assigns a depth factor based on the type of behavior. This algorithmic mechanism ensures that the user profile vector built by the system reflects the user's most authentic technical concerns in real time, rather than being obscured by outdated historical data.

[0063] The semantic parsing module undertakes the crucial tasks of data cleaning and structuring. Addressing the characteristics of lengthy patent documents and obscure legal terminology, this module utilizes a finely tuned large language model to perform a specific extraction task: extracting structured triplet information of **"technical problem-technical means-technical effect"**. This module transforms the unstructured full text into logically rigorous knowledge blocks, providing high-quality input for subsequent vectorization and accurate factual basis for generating briefings, effectively avoiding the illusion problem of model-generated content.

[0064] The intelligent matching module is the system's decision-making hub, its core advantage lying in the introduction of a multi-dimensional value assessment system. In addition to conventional semantic similarity calculation, this module integrates two advanced algorithms: first, a K-Means clustering algorithm optimized based on the elbow rule, which dynamically identifies the user's historical reading interest boundaries to accurately calculate novelty indicators, helping users discover cross-disciplinary technologies; second, a K-Shell decomposition algorithm, which analyzes complex patent citation networks to accurately locate core competitor nodes, thereby quantifying competitive risk indicators. Finally, the module fuses these indicators using a multi-factor linear weighted summation formula, outputting a unique comprehensive score to ensure that the recommendation results not only match interests but also possess insightful and strategic early warning value.

[0065] The briefing generation module achieves personalized information presentation. The core of this module lies in the application of a role-based attention masking mechanism. Depending on whether the user is in R&D, decision-making, or legal roles, the module activates different mask matrices during the large language model generation process, forcing the model to focus on specific parts of the triples (e.g., R&D focuses on technical means, legal focuses on claims). This mechanism significantly reduces the barrier to entry and time cost for users with different backgrounds to read patent information.

[0066] Finally, the feedback optimization module endows the system with the ability to self-evolve. This module first uses a feedback reward value accumulation formula to transform discrete user behaviors such as clicks, favorites, or ignores into numerical reward signals that the computer can understand. Then, the module runs a policy gradient algorithm, a reinforcement learning method, which adjusts the weight parameters of each factor in the intelligent matching module in reverse based on the reward signals. Through this closed-loop mechanism, the system can automatically adapt to changes in user preferences during operation, such as automatically increasing the novelty weight of recommendations for highly active users, thereby achieving continuous optimization of the push strategy.

[0067] Example 3 Further explanation is provided in conjunction with Examples 1-2, such as Figure 1-2As shown, the actual deployment of the patent technology briefing intelligent push system based on user profiles and large models described in this invention first requires building a high-performance computing and storage infrastructure. An NVIDIA A100 or H800 GPU server cluster is used as the computing power foundation to support the fine-tuning and inference of the large language model. At the operating system level, Ubuntu Linux is selected, and a Kubernetes cluster (commonly known as a K8s cluster) is deployed for containerized orchestration and management. At the data storage and processing layer, a distributed vector database, Milvus or Faiss, needs to be deployed to store the patent value vectors and user profile vectors generated by the large model encoding, to meet the needs of efficient similarity retrieval of massive amounts of data. Simultaneously, a graph database, Neo4j, is deployed to construct a patent citation network and competitive relationship graph. Using Neo4j's Graph DataScience library (GDS library), the K-Shell decomposition algorithm is run directly within the database to calculate the kernel values ​​and network influence coefficients of competitor nodes. For unstructured full-text patents and structured triplet data, Elasticsearch or MongoDB is used for persistent storage.

[0068] Next, the core algorithm modules were deployed and the models were trained and fine-tuned. The PyTorch deep learning framework was used to load open-source base models such as Llama 3 or Qwen-72B. A patent corpus containing annotations of technical problems, methods, and effects was used, and instruction fine-tuning (SFT) was performed using a technique called Low-Rank Adaptive Relationship (LoRA) to enable accurate extraction of structured information. The fine-tuned model was then deployed as a high-concurrency API service using vLLM or a TGI inference framework called Text Generation Inference. For the user profile building module, a data stream processing service was written in Python, and the NumPy library was used to calculate the mathematical formulas for time decay and interaction depth weighting, updating user vectors in real time. For novelty index calculation, the Scikit-learn machine learning library was integrated, and K-Means clustering and elbow rule operations were periodically performed on historical reading data to determine the optimal number of clusters. and the centroid vector.

[0069] Subsequently, at the business logic layer, a microservice for intelligent matching and briefing generation is developed based on the FastAPI or Flask framework. This service connects the various components through the LangChain orchestration framework: First, it calls the user profile service to obtain the current weight vector, combines the search results from Milvus, the novelty distance calculated by Scikit-learn, and the risk coefficient provided by Neo4j, and calculates the comprehensive score using a multi-factor linear weighted formula. And complete the sorting; in the briefing generation stage, the role attention masking mechanism described in the claims is actually at the prompt engineering level, that is, at the prompt engineering level, designing system prompt templates for different roles, or intervening in the token probability of specific keywords, including claims and embodiment parameters, through the Logit Processor in the decoding stage, forcing the model to focus on specific triple fields for rewriting and generation.

[0070] Finally, the closed-loop feedback optimization module and the front-end interface are deployed. Redis is used to cache real-time user feedback behaviors, including clicks, favorites, and ignores. After data accumulation, it is batch-transmitted to the reinforcement learning training module. This module runs a lightweight policy network, using Ray RLLib or a custom PyTorch script to execute the policy gradient algorithm, based on the feedback reward value. Calculate the gradient and update the parameters of the recommendation weights The updated parameters are then pushed to the matching engine to take effect. The front-end interface is developed using Vue.js or React frameworks, interacts with the back-end through RESTful APIs, displays personalized briefings for different roles to users, and collects interaction data through event tracking, thereby completing the closed-loop deployment of the entire system.

[0071] The above embodiments are merely preferred technical solutions of the present invention and should not be considered as limitations on the present invention. The scope of protection of the present invention should be limited to the technical solutions described in the claims, including equivalent substitutions of the technical features described in the claims. That is, equivalent substitutions and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. A method for intelligently pushing patent technology briefings based on user profiles and large models, characterized by: include: S1. Obtain explicit user identity information and implicit interaction behavior data, and dynamically weight the user's technical interest characteristics based on the time proximity and interaction depth of the user behavior to obtain a user profile vector containing user role tags and technical interest distribution. S2. Using a large language model fine-tuned with domain knowledge, semantically decompose the patent documents to be pushed, extract the structured triple information of "technical problem, technical means, and technical effect" that can represent the essence of the patent technology, as well as the core features of the claims, and map the extracted information into a patent value vector; S3. Calculate the semantic similarity between the user profile vector and the patent value vector, and introduce the technical novelty index and the competition risk index to weight and correct the semantic similarity. Based on the weighted comprehensive score, determine the target patent to be pushed. S4. Using the structured triplet information of the target patent as the generation basis, based on the user role label in the user profile vector, call the corresponding role attention mask mechanism, use the large language model to differentiate and rewrite the generation basis, generate a personalized technical briefing adapted to the user role and push it. S5. Collect user feedback data on the push notifications and dynamically adjust the weight parameters of the weighted correction in S3 using reinforcement learning algorithms.

2. The intelligent push method for patent technology briefings based on user profiles and large models according to claim 1, characterized in that: The dynamic weighting of user technical interest features in step S1 is implemented based on a time decay and interaction depth weighting algorithm. The steps are as follows: S11. Preset a basic weight system, setting download and bookmark behaviors as high weight levels and browsing and searching behaviors as low weight levels to determine the basic weight of a single interaction behavior; S12. Obtain the timestamp of the interaction, calculate the time difference between the timestamp and the current time, and use the exponential decay logic to calculate the time decay coefficient corresponding to the time difference, so that the closer the behavior is to the current time, the larger the retention coefficient. S13. Determine the depth adjustment factor based on the interaction depth level, correct the basic weight, and multiply the corrected basic weight by the time decay coefficient to obtain the effective contribution value of a single behavior. S14. Based on the above steps, calculate the final weight of the specific technical topic, using the following formula: ； in, This indicates the final weight of the technical topic. The base weight representing a single interaction behavior. This represents the time difference between the time the interaction occurred and the current time. This represents the preset time decay coefficient. Indicates the numerical value of the interaction depth level. This represents the depth adjustment factor.

3. The intelligent push method for patent technology briefings based on user profiles and large models according to claim 1, characterized in that: The specific algorithm steps for extracting structured triple information in step S2 include: S21. Utilize a large language model to identify the text structure of the patent specification, and locate the background technology paragraph, the specific implementation method paragraph, and the beneficial effect paragraph through paragraph semantic analysis; S22. Perform entity relation extraction, extract the defect description of the prior art from the background technology paragraph and mark it as "technical problem"; The core technical features and parameters for solving the technical problem are extracted from the independent claims and specific embodiments and marked as "technical means"; Extract performance improvement indicators from the beneficial effects paragraphs and experimental data tables and label them as "technical effects"; S23. Perform logical serialization, concatenating "technical issues," "technical means," and "technical effects" according to logical causal relationships to form a serialized generation base, which serves as the context source for the generation of the presentation.

4. The intelligent push method for patent technology briefings based on user profiles and large models according to claim 1, characterized in that: The method for calculating the technological novelty index in step S3 is implemented by introducing a K-Means clustering algorithm optimized by the elbow rule. The steps include: S31. Obtain the set of patents read in the user's history, map each patent to a high-dimensional feature vector, and construct a historical interest sample set; S32. Set the number of clusters The traversal range, for each Perform pre-clustering, calculate the corresponding intra-cluster sum of squared errors (SSE), and plot. The curve is analyzed and its inflection points are identified. The corresponding inflection points are then assigned to the curves. The value is determined as the optimal number of clusters. ; S33. Employ the K-Means++ algorithm strategy to select samples from the historical interest sample set according to probability. We have an initial centroid, ensuring that the distance between the initial centroids is maximized. S34. Calculate the Euclidean distance from each sample to each centroid, assign the sample to the nearest cluster, and recalculate the mean of each cluster as the new centroid. Repeat this process until the centroid positions converge, and output the final result. Centroid vectors of historical interest clusters; S35. Calculate the Euclidean distance between the patent value vector of the patent to be pushed and the centroid vector of the nearest historical interest cluster to the patent, and use it as the novelty distance; S36. The novelty distance is normalized using the Sigmoid function to calculate the technical novelty index. The calculation formula is as follows: ； in, Indicators representing technological novelty This represents the patent value vector of the patent to be pushed. This represents the centroid vector of the historical interest cluster closest to the patent. This represents the Euclidean distance operation. This represents the preset novelty perception threshold. This represents the adjustment coefficient.

5. The intelligent push method for patent technology briefings based on user profiles and large models according to claim 1, characterized in that: The method for calculating the competition risk index in step S3 is implemented by introducing the K-Shell decomposition algorithm, and the steps include: S37. Construct a large-scale patent citation network graph that includes competitor nodes of interest to users and a massive number of patent nodes; S38. Perform K-Shell decomposition on the referenced network graph, iteratively removing nodes with a degree less than $ks$, until the degree of all remaining nodes in the network is not less than $ks$. Define the number of levels removed as the kernel value of the node. The network influence coefficient is determined based on the core values ​​of competitor nodes. ; S39. If the applicant of the patent to be pushed is directly in the competitor list, assign the highest risk value; if not, use the breadth-first search algorithm to calculate the minimum path hops between the patent node to be pushed and the competitor node. S310. The competition risk index is calculated by combining the network influence coefficient and the number of path hops. The calculation formula is as follows: ； in, Indicators of competitive risk This indicates the applicant whose patent is to be submitted. This indicates a list of competitor entities. This represents the competitor's network influence coefficient obtained based on K-Shell decomposition. This represents the minimum path hop count. This indicates the maximum number of risk warning jumps.

6. The intelligent push method for patent technology briefings based on user profiles and large models according to claim 1, characterized in that: Step S3 determines the target patents to be pushed based on a comprehensive score, and the calculation of this comprehensive score includes the following sub-steps: S311. Obtain the calculated semantic similarity, technological novelty index, and competition risk index respectively; S312. The three indicators above are linearly weighted and summed using preset or dynamically adjusted weight parameters to obtain a comprehensive score; S313. The formula for calculating the overall score is: ； in, This represents the overall score. The cosine similarity between the user profile vector and the patent value vector is represented. Indicators representing technological novelty Indicators of competitive risk This represents the corresponding weight parameter.

7. The intelligent push method for patent technology briefings based on user profiles and large models according to claim 1, characterized in that: The steps in step S4 that call the corresponding character attention mask mechanism include: S41. If the user role label is R&D engineer, activate the attention mask of the technical implementation dimension. When generating the large language model, assign high attention weights to the information marked as "technical means" and "specific implementation method" in the generation basis, and suppress the output of other information. S42. If the user role label is corporate decision-maker, activate the attention mask of the market value dimension. When generating the large language model, assign high attention weights to information marked as "technical effect" and "technical problem" in the generation base, and associate it with the patent family layout information. S43. If the user role is labeled as intellectual property legal counsel, activate the attention mask for the legal protection dimension and assign high attention weights to information marked as "core features of claims" in the generation base during the generation of the large language model.

8. The intelligent push method for patent technology briefings based on user profiles and large models according to claim 1, characterized in that: Step S5, which involves collecting feedback behavior data and calculating feedback reward values, includes: S51. Capture user actions on a single briefing and categorize these actions as clicks, favorites, forwards, ignores, and negative feedback; S52. Assign corresponding reward coefficients to different types of operation behaviors, setting click behavior as a positive small value, collection and forwarding behavior as a positive large value, and ignoring and negative feedback behavior as a negative value; S53. The formula for calculating the feedback reward value by summing up the rewards generated by all operations is as follows: ； in, This indicates the feedback reward value. This represents the total number of operations. Indicates the first The occurrence state variable of each operation behavior. This represents the reward coefficient for the corresponding action.

9. The intelligent push method for patent technology briefings based on user profiles and large models according to claim 8, characterized in that: The dynamic adjustment of weight parameters in step S5 is implemented using the policy gradient algorithm, and the steps include: S54. Construct a policy network, with user profile vectors as input states and weight parameter combinations as action outputs; S55. Based on the calculated feedback reward value, update the parameters of the policy network using the gradient ascent method to increase the probability of high reward weight combinations being selected; S56. The formula for updating the weight parameters is: ； in, Represents the set of control weight parameters in the policy network. Network parameters, Indicates the learning rate. Indicates the current user profile status. Select a specific weight combination The probability, This indicates the feedback reward value.

10. A patent technology briefing intelligent push system based on user profiles and large models, characterized by: include: The profile building module is used to acquire user data, execute the time decay and interaction depth weighting algorithm in claim 2, and build a user profile vector. The semantic parsing module is used to perform the steps in claim 3, extracting the structured triplet information of the patent using a large language model and quantizing it; The intelligent matching module is used to calculate semantic similarity, calculate the novelty index of technology by combining the K-Means clustering algorithm optimized by the elbow rule in claim 4, calculate the competition risk index by combining the K-Shell decomposition algorithm in claim 5, and calculate the comprehensive score and screen target patents using the formula in claim 6. The briefing generation module is used to execute the role attention masking mechanism in claim 7 and generate personalized technical briefings using a large language model. The feedback optimization module is used to collect user feedback and dynamically update the weight parameters in the intelligent matching module using the formulas and algorithms in claims 8 and 9.

Images