Self-evolution patent emerging technology discovery method and device

Abstract

The embodiment of the invention discloses a self-evolution patent emerging technology discovery method and device, and belongs to the technical field of artificial intelligence and patent information analysis crossing. The method comprises the following steps: receiving a newly input patent text, and carrying out semantic embedding processing on the patent text to generate a fusion vector; based on the similarity with each historical technical cluster center and a dynamic threshold value calculated according to the cluster activeness, the fusion vectors lower than the dynamic threshold value are classified into a candidate new cluster pool, and the candidate new cluster pool is used for storing fusion vectors which cannot be classified into any existing technical cluster; when the candidate new cluster pool meets a multi-dimensional triggering condition, calling a large language model to perform multi-task deep analysis on patents in the pool, and generating an emerging technology early warning report; pushing the emerging technology early warning report to a target user terminal, and receiving target user feedback; and according to the feedback of the target user, adjusting the clustering parameters and the analysis capability of the large language model, updating the clustering parameters required for calculating the dynamic threshold and optimizing the model parameters of the large language model.

Description

Technical Field

[0001] This application relates to the interdisciplinary field of artificial intelligence and patent information analysis, and more specifically, to a self-evolving method and apparatus for discovering emerging patent technologies. Background Technology

[0002] With increasingly active technological innovation activities and a continuous and rapid increase in the number of patent documents, higher demands are being placed on efficient and intelligent patent classification and emerging technology identification methods. Currently, technical solutions based on unsupervised clustering algorithms such as KMeans and hierarchical clustering combined with text embedding models such as BERT and BGE have been widely used in patent text clustering. However, mainstream methods mostly adopt a batch processing mode, re-clustering the entire dataset only after a certain scale of newly added patents has been accumulated. This results in high computational costs, high response latency, and difficulty in timely reflecting the dynamic evolution of the technology field. Although some systems have introduced large language models to enhance semantic understanding capabilities, they still rely on batch processing architectures and mainly focus on summarizing established hot technologies. They lack effective accumulation and identification mechanisms for emerging technology "trailblazers" that are supported by only a few patents and whose semantic features are not yet significant. In addition, the tags or summaries generated by existing solutions are mostly descriptive content, rarely providing forward-looking trend predictions or classification suggestions for examination practice, thus limiting their supporting role in patent examination and science and technology decision-making. Summary of the Invention

[0003] This application provides a self-evolving method and apparatus for discovering emerging patent technologies to address problems existing in the prior art.

[0004] According to a first aspect of the embodiments of this application, a self-evolving method for discovering emerging patent technologies is provided, comprising: Receive new input patent text, perform semantic embedding processing on the patent text, and generate a fusion vector; Based on the similarity with each historical technology cluster center and a dynamic threshold calculated according to cluster activity, fusion vectors below the dynamic threshold are assigned to a candidate new cluster pool, which is used to store fusion vectors that cannot be assigned to any existing technology cluster. When the candidate new cluster pool meets the multi-dimensional triggering conditions, the large language model is invoked to perform multi-task in-depth analysis of the patents in the pool and generate an emerging technology early warning report. The emerging technology early warning report is pushed to the target user's terminal, and feedback from the target user is received; Based on the feedback from the target users, the clustering parameters and the large language model analysis capabilities are adjusted, the clustering parameters required to calculate the dynamic threshold are updated, and the model parameters of the large language model are optimized, thereby achieving the self-evolution of the system.

[0005] This solution constructs a complete workflow integrating real-time processing, early warning detection, intelligent analysis, and self-evolution. First, semantic embedding transforms patent text into high-quality fusion vectors, providing a precise semantic foundation for subsequent analysis. Second, an incremental clustering mechanism based on dynamic thresholds avoids full re-clustering while enabling real-time classification of new patents, temporarily storing unclassifiable patents in a candidate cluster pool, effectively accumulating potential emerging technology signals. Third, multi-dimensional condition-triggered deep analysis using a large language model enables intelligent assessment of patents within the pool, generating emerging technology early warning reports that include technology naming, trend prediction, and classification suggestions, directly supporting examination decisions. Finally, a user feedback loop is introduced, continuously optimizing clustering parameters and model capabilities using expert knowledge, enabling the system to self-learn and evolve. This solution fundamentally solves the technical challenges of low computational efficiency, inability to respond in real-time, and lack of forward-looking perception and deep decision support in existing technologies, achieving dynamic, efficient, and intelligent evolution of the patent classification system under continuous data flow.

[0006] In one possible implementation, the semantic embedding process of the patent text to generate a fusion vector includes: Extract the title, abstract, and claims from the patent text. The title, abstract, and claims are vectorized using a pre-trained embedding model to obtain three independent semantic vectors. The technical field to which the patent belongs is determined based on the patent's classification number or text content; Query the preset field-weight mapping table to obtain the title weight, abstract weight, and claim weight of the corresponding field. The three independent semantic vectors are weighted and summed according to the weights to generate a preliminary fusion vector; After encoding the external meta-attributes of the patent, they are fused with the preliminary fusion vector to form an enhanced fusion vector; The enhanced fusion vector is normalized, and the output is the final fusion vector.

[0007] This paper proposes a patent domain-aware, dynamically weighted semantic vector generation method. By extracting and encoding the three core text components—title, abstract, and claims—the independent semantic information of each is preserved. Furthermore, based on the patent classification number or content, the technical field of the patent is automatically identified, and a pre-defined domain-weight mapping table is consulted to adaptively assign differentiated fusion weights to patents in different domains (e.g., chemical materials patents emphasize claims, software algorithm patents emphasize abstracts, and electronic power patents increase title weight). This makes the generated preliminary fusion vectors more accurately reflect the core innovation of patents in different domains. Finally, by introducing and integrating external meta-attributes of the patent (such as citation count and applicant type), the vector's ability to represent dimensions such as technological influence and innovation subject characteristics is enhanced. Normalization ensures the stability of the vectors in similarity calculation. The quality of the fusion vectors generated by this method is significantly better than simple text averaging or static weighting methods, providing a reliable input foundation for subsequent high-precision clustering and discovery.

[0008] In one possible implementation, the external meta-attribute includes at least one of the following: number of citations, applicant type, and number of patent families.

[0009] This scheme clarifies the specific composition of external meta-attributes, including dimensions such as citation count, applicant type, and number of patent families. These attributes are important indicators for measuring the influence of patented technologies, the activity level of innovation entities, and the breadth of technological layout. Encoding these attributes and fusing them with textual semantic vectors results in enhanced fusion vectors that not only contain deep technical semantic information but also rich non-textual contextual features. For example, the "influence" dimension of highly cited patents is strengthened in the vector space, helping to identify key technological foundations; differences in applicant types between universities and enterprises may imply different R&D models and innovation priorities. This multi-source information fusion significantly improves the discriminative power of semantic vectors and the accuracy of clustering in representing the real technological landscape, enabling the system to understand each patent more comprehensively and holistically.

[0010] In one possible implementation, the step of assigning fusion vectors below a dynamic threshold to the candidate new cluster pool based on similarity to each historical technology cluster center and a dynamic threshold calculated according to cluster activity includes: Calculate the similarity between the fusion vector and the center of each historical technology cluster, and determine the target historical technology cluster with the highest similarity to the fusion vector; Maintain an activity index for each historical technology cluster that is positively correlated with the number of patents contained in that cluster; Based on the preset base threshold, adjustment coefficient and the activity index, a dynamic affiliation threshold is calculated for each cluster; The fusion vector is compared with the dynamic attribution threshold corresponding to the technology cluster with the highest similarity. If the similarity of the fusion vector is lower than the dynamic attribution threshold, the fusion vector is assigned to the candidate new cluster pool. If the similarity of the fusion vector is not lower than the dynamic attribution threshold, the fusion vector is added to the technology cluster with the highest similarity, and the center of the technology cluster is updated by using a moving average method.

[0011] This scheme implements an intelligent and adaptive incremental clustering and triage mechanism. First, it calculates the similarity between new patents and all historical cluster centers, providing a basis for classification. Second, it innovatively introduces an "activity" index positively correlated with the number of patents contained in each cluster, and calculates a dynamic classification threshold based on this index, a base threshold, and an adjustment coefficient. This allows mature, stable large clusters to use higher thresholds to maintain classification rigor, while emerging, small clusters with fewer patents use lower thresholds to reduce the risk of "false negatives" (i.e., avoiding misclassifying patents belonging to the cluster as unclassifiable), significantly improving the flexibility and accuracy of classification. Finally, for classifiable patents, a moving average method is used to locally update cluster centers, achieving a smooth evolution of cluster representation and avoiding drastic fluctuations in centers caused by single patents. This mechanism, while ensuring high processing efficiency, greatly enhances the system's adaptability to dynamic technological evolution (including the growth and evolution of existing clusters).

[0012] In one possible implementation, updating the center of the technology cluster using a moving average method involves balancing the weights of the historical center of the technology cluster with those of the new patent vector using an adjustable decay factor.

[0013] This scheme introduces an adjustable decay factor into the cluster center update process. This factor balances the weight of historical cluster center information and newly added patent vectors when calculating new centers. A higher decay factor means greater reliance on historical centers, resulting in more conservative and stable updates, which is beneficial for maintaining the stability of mature technology clusters; a lower decay factor means that new patent vectors have a greater impact, resulting in more sensitive updates, which is beneficial for emerging clusters to quickly form their representative centers. By systematically adjusting this factor, differentiated update strategies can be adopted for technology clusters at different development stages (stable period or rapid growth period) or with different characteristics (such as susceptibility to abnormal patent interference), thereby optimizing the overall stability of clustering results and the speed of adaptation to new data, and further improving the robustness of the incremental clustering system.

[0014] In one possible implementation, when the candidate new cluster pool meets the multi-dimensional triggering conditions, a large language model is invoked to perform multi-task deep analysis of the patents within the pool, generating an emerging technology early warning report, including: Monitor whether the growth in the number of patents in the candidate new cluster pool reaches a preset condition; Analyze whether the novelty of technical terms in the patents within the candidate new cluster pool exceeds a preset threshold; Whether the geographical distribution of patent applicants within the candidate new cluster pool meets a preset standard; When any one of the three dimensions of quantity growth, terminology novelty, and geographical distribution diffusion is met, a representative set of patents is selected from the candidate new cluster pool. Construct structured prompt templates to guide the large language model to perform three collaborative tasks: naming technology directions, predicting technology trends, and providing classification suggestions. The outputs of the three tasks are integrated into a structured early warning report on emerging technologies.

[0015] This solution designs a precise and reliable mechanism for triggering and deeply analyzing emerging technology trends. It abandons a single quantitative threshold and employs cross-monitoring across three dimensions: quantity growth, novelty of technical terminology, and geographical diffusion of applicants. Triggering is achieved when any one dimension meets the criteria, ensuring trigger sensitivity (not missing early signals) while reducing false alarm rates through multi-dimensional verification (e.g., a large quantity of patents with outdated terminology may simply be a continuation of existing technology). After triggering, a representative patent set is selected, and a structured prompt template is constructed to systematically guide the large language model to perform three collaborative tasks: naming the technology direction, predicting trends, and suggesting classifications. This multi-task parallel analysis can deeply mine the commonalities and future value of the patent set from different perspectives, integrating the results into a structured early warning report. This mechanism enables the system not only to discover "hot spots" but also to proactively identify and analyze disruptive technology directions that are still in their nascent stages but possess development potential.

[0016] In one possible implementation, the technology direction naming task is used to generate concise technology labels; the technology trend prediction task is used to output future technology development trends; and the classification suggestion task is used to output relevant patent classification numbers.

[0017] This solution clarifies the specific outputs of the three core tasks of the large language model. The technology direction naming task generates concise and accurate technology tags, helping target users quickly grasp the essence of the technology; the technology trend prediction task outputs an assessment of key technology developments over the next 1-2 years, providing a forward-looking perspective and assisting in the early planning of examination strategies; the classification suggestion task outputs the most relevant patent classification numbers (such as IPC / CPC), providing direct reference for accurate and rapid patent classification. These three tasks support each other, jointly forming a comprehensive and directly usable intelligent report to support decision-making. Compared to existing technologies that can only provide descriptive summaries or simple tags, the report generated by this invention has stronger insight, predictive power, and operability, significantly improving the intelligence level and decision support capabilities of patent examination and intelligence analysis.

[0018] In one possible implementation, adjusting clustering parameters and large language model analysis capabilities based on the target user feedback, updating the clustering parameters required to calculate the dynamic threshold, and optimizing the model parameters of the large language model to achieve system self-evolution includes: Record the type of action the target user takes in response to the emerging technology warning report, including confirmation, rejection, or modification. If the operation type is rejection, then the common features of the corresponding batch of patents are extracted and used to raise the threshold of the relevant triggering conditions. If the operation type is confirmation or modification, the technical information finally adopted by the target user will be used as a supervision tag. A training dataset is constructed using all feedback data, and a parameter-efficient fine-tuning method is used to optimize the large language model; Simultaneously, based on feedback information, the adjustment coefficient and the attenuation factor are dynamically adjusted.

[0019] This solution establishes a highly efficient human-machine collaborative closed-loop self-evolutionary mechanism. The system records the target user's specific actions on warning reports (confirmation, rejection, or modification) and transforms this feedback into "fuel" for system optimization. For rejected reports, the system automatically analyzes their common characteristics and raises the relevant trigger condition thresholds, thereby reducing similar false positives in the future. For confirmed or modified reports, the correct information ultimately adopted by the target user (such as accurate technical names and classification numbers) is used as high-quality supervision labels. A training set is constructed using all feedback data, and efficient parameter fine-tuning techniques (such as LoRA) are employed to continuously optimize the large language model, bringing its naming, prediction, and classification suggestion capabilities closer to expert levels. Simultaneously, feedback information is also used to dynamically adjust key parameters in the clustering process (such as adjustment coefficients and decay factors), making the clustering behavior more consistent with expert domain knowledge. This mechanism ensures that the system can "learn from practice," becoming more accurate with use and achieving true intelligent evolution.

[0020] In one possible implementation, the initial values ​​of the adjustment coefficient and the attenuation factor are determined by an optimization algorithm during the system offline phase.

[0021] This solution provides a scientific and systematic method for initializing key system parameters (adjustment coefficients, attenuation factors). Before offline deployment, historical patent datasets are used to automatically search for optimal initial values ​​of the parameters, using algorithms such as Bayesian optimization with clustering accuracy or stability as optimization objectives. This avoids the subjectivity and suboptimal nature that may result from relying on manual experience, ensuring a high-performance starting point after system deployment. Combined with an online dynamic adjustment mechanism, a complete parameter management system of "offline optimization initialization and online feedback adaptation" is formed, ensuring that the system maintains optimal or near-optimal performance throughout its entire lifecycle, thus improving the reliability and engineering applicability of the method.

[0022] According to a second aspect of the embodiments of this application, a self-evolving patent emerging technology discovery device is provided, comprising: The receiving unit is used to receive new input patent text, perform semantic embedding processing on the patent text, and generate a fusion vector; The processing unit is used to classify fusion vectors that are below the dynamic threshold into the candidate new cluster pool based on the similarity with each historical technology cluster center and the dynamic threshold calculated according to the cluster activity. The candidate new cluster pool is used to store fusion vectors that cannot be classified into any existing technology cluster. The generation unit is used to call a large language model to perform multi-task in-depth analysis of the patents in the pool when the candidate new cluster pool meets the multi-dimensional triggering conditions, and generate an emerging technology early warning report. The sending unit is used to push early warning reports on emerging technologies to target user terminals and receive feedback from target users; The adjustment unit is used to adjust the clustering parameters and the large language model analysis capabilities based on the feedback from the target users, update the clustering parameters required to calculate the dynamic threshold and optimize the model parameters of the large language model, thereby realizing the self-evolution of the system.

[0023] According to a third aspect of the embodiments of this application, the present invention provides an electronic device, including: a memory and a processor, the memory and the processor being communicatively connected to each other, the memory storing computer instructions, and the processor executing the computer instructions to perform the patent emerging technology discovery method of clustering self-evolution based on large language model evaluation feedback described in the first aspect or any corresponding embodiment.

[0024] According to a fourth aspect of the embodiments of this application, the present specification provides a computer-readable storage medium storing computer instructions that, when executed by a processor, implement the self-evolving patent emerging technology discovery method as described in any of the preceding claims.

[0025] According to a fifth aspect of the embodiments of this application, this specification provides a computer program product or computer program, the computer program product including a computer program stored in a computer-readable storage medium; a processor of a computer device reads the computer program from the computer-readable storage medium, and the processor executes the computer program to implement the self-evolving patent emerging technology discovery method as described in any of the preceding claims.

[0026] The technical effects achieved by the third to fifth aspects mentioned above are similar to those achieved by the corresponding technical means in the first aspect, and will not be repeated here. Attached Figure Description

[0027] Figure 1 This is a flowchart illustrating the self-evolving patent emerging technology discovery method provided in the embodiments of this application; Figure 2 This is a schematic diagram of the structure of the self-evolving patent emerging technology discovery device provided in the embodiments of this application; Figure 3 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0028] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0029] It should be understood that "multiple" as mentioned herein refers to two or more. In the description of the embodiments of this application, unless otherwise stated, " / " means "or," for example, A / B can mean A or B; "and / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. In addition, to facilitate a clear description of the technical solutions of the embodiments of this application, the terms "first," "second," etc., are used in the embodiments of this application to distinguish identical or similar items with substantially the same function and effect. Those skilled in the art will understand that the terms "first," "second," etc., do not limit the quantity or execution order, and the terms "first," "second," etc., do not necessarily imply that they are different.

[0030] Furthermore, the terms “comprising” and “having”, and any variations thereof, are intended to cover non-exclusive inclusion, such that a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such process, method, product, or apparatus.

[0031] As mentioned in the background section, with increasingly active scientific and technological innovation activities, the number of patent documents continues to grow rapidly, placing higher demands on efficient and intelligent patent classification and emerging technology identification methods. Currently, technical solutions based on unsupervised clustering algorithms such as KMeans and hierarchical clustering combined with text embedding models such as BERT and BGE have been widely used in patent text clustering. However, mainstream methods mostly adopt a batch processing mode, re-clustering the entire dataset only after a certain scale of newly added patents has been accumulated, resulting in high computational overhead, high response latency, and difficulty in timely reflecting the dynamic evolution of the technology field. Although some systems introduce large language models to enhance semantic understanding capabilities, they still rely on batch processing architectures, mainly focusing on summarizing hot technologies that have already reached a certain scale, lacking an effective accumulation and identification mechanism for emerging technology "seeds" that are supported by only a few patents and whose semantic features are not yet significant. In addition, the tags or summaries generated by existing solutions are mostly descriptive content, rarely providing forward-looking trend predictions or classification suggestions for examination practice, limiting their supporting role in patent examination and science and technology decision-making.

[0032] This invention provides a self-evolving method for discovering emerging patent technologies, and, as such... Figure 1 As shown, it includes the following steps: S101: Receive the new input patent text, perform semantic embedding processing on the patent text, and generate a fusion vector.

[0033] In practice, the first step is to acquire newly input patent text data, including structured or unstructured text content such as titles, abstracts, and claims. To achieve high-quality semantic representation, this step employs a domain-aware, dynamically weighted semantic vector generation method.

[0034] First, the title, abstract, and claims are extracted from the patent text. Then, pre-trained embedding models (such as Qwen3-Embedding and bge-large-zh-v1.5) are used to vectorize the three parts of the text, resulting in three independent semantic vectors.

[0035] Furthermore, the patent's technical field is automatically identified based on its International Patent Classification (CPC), for example, by mapping according to the CPC main classification number (e.g., G06F for the computer field, A01 for the agricultural field). Fine-grained identification of the technical field can also be achieved by combining keywords in the patent abstract or through a pre-trained field classification model to improve accuracy. The preset field-weight mapping table is shown below.

[0036]

[0037] For example, for chemical and materials patents, the core lies in the specific compound structure, proportions, and synthesis routes, which are concentrated in the claims. Therefore, the claim weight is set at 0.6, while the title and abstract have lower weights. For software and algorithm patents, the innovation points are usually described more completely in the abstract, so the abstract weight is increased to 0.4. For electronics and power patents, the title itself contains information about key components, so the title weight is set at 0.25, significantly higher than other fields. This field-adaptive weighting method allows the generated semantic vectors to more accurately reflect the core innovation points of patents in different technical fields.

[0038] Then, the dynamic fusion weights of the title, abstract, and claims in the corresponding technical field are obtained. For example, patents in the field of chemical materials place more emphasis on the claims, so their claim weights are set to 0.6, while the title and abstract weights are 0.1 and 0.3 respectively; for patents in the field of software algorithms, the abstract weight is increased to 0.4.

[0039] The three independent semantic vectors are weighted and summed to generate a preliminary fusion vector. To enhance the vector's representational power, the patent's external meta-attributes (such as citation count, applicant type (company / university / individual), and number of patent families) are encoded using a lightweight neural network (such as a single-layer MLP) and concatenated with the preliminary fusion vector to form an enhanced fusion vector. For example, a high citation count can strengthen the patent's "influence" dimension in the clustering space, which helps identify key technological points.

[0040] It should be noted that, as another common implementation method, the system can also forgo domain recognition and instead directly use preset fixed weights to weight and fuse the text vectors of each part. Based on experimental experience, the preferred weight allocation is: title weight 0.5, abstract weight 0.3, and claims weight 0.2. In this case, the calculation formula for the fused vector simplifies to: V final = 0.5 V title + 0.3 V abstract + 0.2 V claims This fixed-weight strategy has lower computational overhead and is suitable for mixed patent stream processing scenarios with extremely high real-time requirements or unclear domain characteristics.

[0041] Finally, L2 normalization is performed on the enhanced fusion vector, and the output is the final fusion vector, which is used for subsequent clustering and similarity calculation.

[0042] S102, based on the similarity with each historical technology cluster center and the dynamic threshold calculated according to the cluster activity, the fusion vectors below the dynamic threshold are included in the candidate new cluster pool.

[0043] In practical implementation, this step implements an intelligent adaptive incremental clustering and triage mechanism. First, it calculates the cosine similarity between the new patent fusion vector and the centers of each historical technology cluster. An activity index is maintained for each historical technology cluster, which is positively correlated with the number of patents within the cluster; for example, it is defined as... , where x is the number of patents within the cluster.

[0044] Based on a preset base threshold (e.g., 0.6), adjustment coefficient and activity index Calculate the dynamic affiliation threshold for each cluster. The adjustment coefficient During the offline training phase, based on historical patent datasets, classification accuracy is optimized using a Bayesian optimization algorithm for automatic search and determination. During system operation, user feedback or automatic evaluation metrics can be used to adjust the system's performance using an exponentially weighted moving average mechanism. Dynamic fine-tuning is performed. This design uses a higher threshold for large clusters with many patents and clear boundaries to maintain classification rigor, while using a lower threshold for emerging clusters with fewer patents to reduce the risk of misclassification.

[0045] The new patent vector is compared with the dynamic threshold corresponding to the technology cluster with the highest similarity: if the similarity is not lower than the threshold, the patent is added to the cluster, and the cluster center is updated using a moving average method. The update formula is as follows: ,in A decay factor (e.g., 0.9) is used to balance the weights of the historical center and the new vector. As the new cluster center, For the old cluster centers, it should be understood that the above moving average update method is essentially an efficient approximation of the weighted average in streaming computing scenarios. In terms of specific mathematical implementation, if system resources allow for storing intra-cluster historical vectors, the weighted average formula can also be used to update the cluster centers: ,in It represents the weight of the i-th vector. This embodiment preferably uses a moving average formula because it eliminates the need to reread historical data; it only requires iteration based on the old center, better meeting the real-time processing requirements of the patent stream.

[0046] The attenuation factor Similarly, a systematic approach was used to determine the initial values: during the offline phase, based on the cluster center stability on the validation set, Bayesian optimization was employed to determine the initial values; during system operation, if frequent cluster assignment errors or drastic center fluctuations were detected, the initial values ​​were automatically reduced. To enhance the preservation of historical information, and conversely, to appropriately increase it. To accelerate the adaptation process; if the similarity is below the threshold, the patent is included in the candidate new cluster pool, which is used to temporarily store patents that cannot be included in the existing clusters, serving as an accumulation area for potential emerging technology signals.

[0047] S103, when the candidate new cluster pool meets the multi-dimensional triggering conditions, the large language model is invoked to perform multi-task in-depth analysis of the patents in the pool and generate an emerging technology early warning report.

[0048] In practice, this step incorporates a multi-dimensional dynamic triggering and deep intelligent judgment mechanism. The system continuously monitors multi-dimensional indicators of the candidate new cluster pool, including: 1. Quantitative Growth Dimension: Monitoring the absolute number of newly added patents over N consecutive days. In a specific embodiment, when the number of newly added patents in the candidate new cluster pool is greater than or equal to 3 over N consecutive days (e.g., N=7), it is considered to meet the quantitative growth condition. Rapid quantitative growth is a direct signal of increased technological activity.

[0049] 2. Semantic Novelty Dimension: Calculate the TF-IDF value of the new term or use the embedding space distance method (calculate the Euclidean distance between the vector of the new term and the center of the existing technology lexicon). When its novelty exceeds a preset threshold and appears in multiple consecutive independent patents, it indicates the existence of a new technological paradigm. In a specific embodiment, when the number of times the new term appears in the patent pool is greater than 2, it is considered to meet the new term frequency requirement.

[0050] 3. Social diffusion dimension: Analyze the changes in entropy values ​​of applicants' geographical distribution. If applications that were originally concentrated in a few regions suddenly become widespread (with a significant increase in entropy), it indicates that the technology is attracting the attention of a wider range of subjects and has the potential for cross-regional diffusion.

[0051] When any of the above dimensions meets the triggering conditions, the system selects a representative set of patents from the pool and calls the large language model to perform multi-task deep analysis. The model is guided to perform the following collaborative tasks using structured prompt templates: Technical direction naming task: Generate concise and accurate technical labels (e.g., "small molecule design based on deep generative models"). A typical prompt template is: "You are a senior patent examination analyst. Based on the common technical features of the following patent abstracts and claims, please extract a technical direction name of no more than 8 Chinese characters, highlighting its core innovation. The patent content is as follows: [Patent 1]; [Patent 2]; [Patent 3]..."

[0052] Technology Trend Forecasting Task: Output the key technology development trends for the next 1-2 years. A typical prompt template is: "Based on the application trends of the above patents over the past 12 months, the main applicant types (enterprises / universities), and changes in high-frequency technical terms, please predict three key technology development trends for this technology direction in the next 1-2 years. Please output in the numbering format of '1. … 2. … 3. …', with each trend not exceeding 30 characters."

[0053] IPC / CPC classification suggestion task: Recommend the most relevant patent classification numbers. A typical prompt template is: "Please recommend the most matching CPC classification numbers (maximum 3) for the above patent set, sorted by relevance from high to low. Output only the classification numbers, in the following format: H04L67 / 1097, G06F21 / 62, H04W12 / 04."

[0054] The outputs of the three tasks are integrated into a structured emerging technology early warning report, which includes technology naming, trend predictions and classification suggestions, and can be directly used for review and decision support.

[0055] S104 pushes early warning reports on emerging technologies to target user terminals and receives feedback from target users.

[0056] In practice, the system pushes the generated warning report to the target user's terminal interface. The target user can be a patent examiner or relevant decision-maker, or any other user who needs to use this content; this embodiment does not limit this. The report includes a system confidence score, generated based on indicators such as trigger strength and model output consistency. Users can perform three types of operations on the report: confirm (accept the report content), reject (believe the analysis is incorrect), and modify / supplement (provide more accurate naming, trend, or classification suggestions). The system records the user's operation type and related feedback information, providing monitoring signals for subsequent self-evolution.

[0057] S105, Based on feedback from target users, adjust clustering parameters and large language model analysis capabilities, update the clustering parameters required for calculating dynamic thresholds, and optimize the model parameters of the large language model.

[0058] In practice, this step constructs a system self-evolution mechanism based on feedback loops, employing a dual-path optimization architecture: 1. Adaptive calibration of clustering parameters: For clusters of "false positives" that are frequently rejected by examiners, the system automatically analyzes their common features (such as specific text patterns), dynamically adjusts the similarity threshold or semantic vector weights (such as reducing the weight of a certain type of text), and optimizes the incremental clustering strategy.

[0059] 2. Continuous Evolution of Large Model Intelligence: The output of the large model and the reviewer's correction results are used to construct a "prediction-real" feedback dataset, which is used for: lightweight fine-tuning of the large model (such as LoRA) to improve the accuracy of naming, prediction and classification suggestions; automatically optimizing the design of prompt words to improve the task guidance effect; and continuously optimizing the "emergence scoring model" through reinforcement learning mechanism.

[0060] Specifically, the system performs the following optimizations based on user feedback: If the report is rejected, extract the common features of the corresponding batch of patents and raise the threshold of the relevant triggering conditions to reduce false alarms of the same type. If the report is confirmed or modified, the technical information finally adopted by the user (such as the correct classification number and technical name) will be used as a high-quality supervision label; A training dataset is constructed using all feedback data, and a parameter-efficient fine-tuning method (such as LoRA) is used to continuously optimize the large language model, thereby improving its naming, prediction, and classification suggestion capabilities. Dynamically adjust key parameters in incremental clustering, such as adjustment coefficients. With decay factor This makes it more in line with the cognitive and cluster stability requirements of domain experts.

[0061] Through the above mechanism, the system achieves a closed loop from "human feedback" to "model and parameter optimization." Specifically, by updating the parameters for calculating dynamic thresholds (such as adjustment coefficients) and fine-tuning the parameters of the large language model, the system achieves its self-evolution capability. This continuously improves classification accuracy, early warning reliability, and decision support capabilities.

[0062] In one example, suppose the system is continuously monitoring a global patent database when a newly published patent enters the processing flow. This patent relates to a novel vaccine for the prevention of respiratory syncytial virus (RSV), whose core technology utilizes messenger RNA (mRNA) technology, encapsulated in specific lipid nanoparticles for delivery. The system first performs semantic embedding processing on this patent. It automatically parses the patent's title, abstract, and claims text, and identifies its International Patent Classification (CPC) as pointing to the pharmaceutical field. Based on preset field weighting rules, the system assigns the highest fusion weight to the claims section, as the core technical details and scope of protection of pharmaceutical patents are usually most rigorously defined in this part; the abstract and title have relatively lower weights. After generating the initial semantic vector, the system also incorporates the patent's meta-information, such as the applicant being an emerging biotechnology company and having a certain number of citations. This information is encoded and fused with the textual semantic vector, ultimately forming a fusion vector that comprehensively represents the patent's technical content and impact.

[0063] Next, the system compares the fusion vector of this new patent with the similarity of existing technology cluster centers in the historical database. The historical database may contain large clusters related to technologies such as "traditional inactivated virus vaccines" or "recombinant protein vaccines." Although the new patent has a certain semantic connection to the broad category of "virus vaccines," the calculated similarity does not reach the higher dynamic threshold set by the system for these mature and stable large clusters. This dynamic threshold is automatically adjusted based on the "activity" of each cluster (related to the number of patents it contains), with larger clusters having a higher threshold to ensure the rigor of classification. Therefore, the system determines that this patent cannot be directly classified into any existing technology category and places it in a buffer pool called the "candidate new cluster pool." This pool is specifically used to accumulate patents that differ from the existing technology framework and may represent emerging trends.

[0064] In the following weeks, the system received several more patents involving similar mRNA technology pathways but targeting different viral targets or employing novel lipid formulations. These were also grouped into the same candidate new cluster pool due to insufficient similarity. The system continuously monitored the status of this pool. Soon, the rapid increase in the number of related patents in the pool triggered preset conditions. Simultaneously, text analysis revealed that the patents in the pool frequently used technical terms uncommon in previous patent databases, such as "ionizable lipids" and "self-replicating sequences," and their novelty indicators exceeded the threshold. Furthermore, the applicants for these patents expanded from one or two institutions initially to research teams and companies in multiple countries. These changes in quantity, terminology, and applicant distribution collectively constituted a multi-dimensional trigger signal.

[0065] Once triggered, the system automatically selects several of the most representative patent documents from the pool and invokes the large language model for in-depth analysis. Through carefully designed instructions, the large language model performs three core tasks: First, it comprehensively analyzes these patents, extracts and generates an accurate technical direction name, such as "a respiratory syncytial virus mRNA vaccine platform based on novel lipid nanoparticle delivery"; second, based on the patent application time series and content evolution, it predicts possible future technological development trends, such as focusing on extrahepatic targeted delivery efficiency or combining novel adjuvant development; third, it provides suggested international patent classification numbers. These outputs are automatically integrated into a structured "Emerging Technology Early Warning Report," along with a confidence score of the system's analysis results.

[0066] This report is then sent to the relevant patent examiner. As a domain expert, the examiner reviews the report, approves its analytical conclusions, and may make minor adjustments to the technology naming or classification number based on the system's suggestions before confirming the adoption of the report. This "confirmation" feedback is crucial to the system. The system uses the technology names and classification numbers finally confirmed by the examiner as high-quality annotation data for subsequent fine-tuning of the large language model, making its future naming and classification suggestions more accurate. Simultaneously, the successfully identified pattern is recorded and used to optimize the trigger threshold of the "candidate new cluster pool" and the similarity judgment parameters for incremental clustering. Thus, this emerging technology direction related to mRNA vaccines is formally incorporated into the system's knowledge system, forming an initial technology cluster for the rapid classification of subsequent new patents. Through this complete process, the system achieves a closed loop from perceiving "technology emerging trends," intelligent analysis, to self-evolution and knowledge growth through expert feedback.

[0067] The above examples illustrate the method embodiments according to this application. The present invention also provides a self-evolving patent emerging technology discovery device. Figure 2 This is a schematic diagram of the structure of a self-evolving patent emerging technology discovery device according to an embodiment of the present invention. (Refer to...) Figure 2 The self-evolving patented emerging technology discovery device 700 includes the following modules.

[0068] The receiving unit 701 is used to receive newly input patent text, perform semantic embedding processing on the patent text, and generate a fusion vector; The processing unit 702 is used to classify fusion vectors below the dynamic threshold into the candidate new cluster pool based on the similarity with each historical technology cluster center and the dynamic threshold calculated according to the cluster activity. The candidate new cluster pool is used to store fusion vectors that cannot be classified into any existing technology cluster. The generation unit 703 is used to call a large language model to perform multi-task in-depth analysis of the patents in the pool when the candidate new cluster pool meets the multi-dimensional triggering conditions, and generate an emerging technology early warning report. The sending unit 704 is used to push the emerging technology early warning report to the target user terminal and receive feedback from the target user; The adjustment unit 705 is used to adjust the clustering parameters and the large language model analysis capability based on the feedback from the target user, update the clustering parameters required to calculate the dynamic threshold and optimize the model parameters of the large language model, thereby realizing the self-evolution of the system.

[0069] This application also provides a computer program product including computer program instructions that, when executed by a processor, cause the processor to perform the steps of the self-evolving patent emerging technology discovery method according to various embodiments of this specification as described in the "Exemplary Methods" section above.

[0070] Computer program products can be written in any combination of one or more programming languages ​​to perform the operations of the embodiments in this specification. The programming languages ​​include object-oriented programming languages ​​such as Java, C++, etc., as well as conventional procedural programming languages ​​such as the "C" language or similar programming languages.

[0071] This application also provides a computer-readable storage medium having a computer program stored thereon, the computer program being executed by a processor of the steps in the self-evolving patent emerging technology discovery method according to various embodiments of this specification as described in the "Exemplary Methods" section above.

[0072] This application also provides an electronic device, including a memory and a processor. The memory stores a self-evolving method for discovering emerging patent technologies, and the processor is used to employ the self-evolving method for discovering emerging patent technologies when executing the self-evolving method.

[0073] Specifically, such as Figure 3 As shown, the electronic device includes a processor 100, at least one communication bus 200, a user interface 300, at least one external communication interface 400, and a memory 500. The communication bus 200 is configured to enable communication between these components. The user interface 300 may include a display screen, and the external communication interface 400 may include standard wired and wireless interfaces. The memory 500 stores a self-evolving method for discovering emerging patent technologies. The processor 100 is used to employ the aforementioned method when executing the self-evolving method for discovering emerging patent technologies stored in the memory 500.

[0074] The descriptions of the above computer program products, computer-readable storage media, and electronic devices are similar to those of the above method embodiments, and have similar beneficial effects. For any technical details not disclosed in the computer program products, computer-readable storage media, and electronic devices of this application, please refer to the descriptions of the method embodiments of this application for understanding.

[0075] The sequence numbers or order of description of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0076] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0077] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0078] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0079] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions. When the computer instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium accessible to a computer, or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., digital versatile disc (DVD)), or a semiconductor medium (e.g., solid state disk (SSD)). It is worth noting that the computer-readable storage medium mentioned in the embodiments of this application can be a non-volatile storage medium; in other words, it can be a non-transient storage medium. It should be noted that the information (including but not limited to user device information, user personal information, etc.), data (including but not limited to data used for analysis, stored data, displayed data, etc.), and signals involved in the embodiments of this application are all authorized by the user or fully authorized by all parties, and the collection, use, and processing of related data must comply with the relevant laws, regulations, and standards of the relevant countries and regions. For example, the scene data of the current frame in the 3D virtual scene involved in the embodiments of this application, the client's device information, and the scene interaction information are all obtained with full authorization.

[0080] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A self-evolving method for discovering emerging patent technologies, characterized in that, include: Receive new input patent text, perform semantic embedding processing on the patent text, and generate a fusion vector; Based on the similarity with each historical technology cluster center and a dynamic threshold calculated according to cluster activity, fusion vectors below the dynamic threshold are assigned to a candidate new cluster pool, which is used to store fusion vectors that cannot be assigned to any existing technology cluster. When the candidate new cluster pool meets the multi-dimensional triggering conditions, the large language model is invoked to perform multi-task in-depth analysis of the patents in the pool and generate an emerging technology early warning report. The emerging technology early warning report is pushed to the target user's terminal, and feedback from the target user is received; Based on the feedback from the target users, update the clustering parameters required to calculate the dynamic threshold and optimize the model parameters of the large language model.

2. The method according to claim 1, characterized in that, The step of semantic embedding the patent text to generate a fusion vector includes: Extract the title, abstract, and claims from the patent text. The title, abstract, and claims are vectorized using a pre-trained embedding model to obtain three independent semantic vectors. The technical field to which the patent belongs is determined based on the patent's classification number or text content; Query the preset field-weight mapping table to obtain the title weight, abstract weight, and claim weight of the corresponding field. The three independent semantic vectors are weighted and summed according to the weights to generate a preliminary fusion vector; After encoding the external meta-attributes of the patent, they are fused with the preliminary fusion vector to form an enhanced fusion vector; The enhanced fusion vector is normalized, and the output is the final fusion vector.

3. The method according to claim 2, characterized in that, The external meta-attributes include at least one of the following: number of citations, applicant type, and number of patent families.

4. The method according to claim 1, characterized in that, The method of incorporating fusion vectors below the dynamic threshold into the candidate new cluster pool based on similarity to each historical technology cluster center and a dynamic threshold calculated according to cluster activity includes: Calculate the similarity between the fusion vector and the center of each historical technology cluster, and determine the target historical technology cluster with the highest similarity to the fusion vector; Maintain an activity index for each historical technology cluster that is positively correlated with the number of patents contained in that cluster; Based on the preset base threshold, adjustment coefficient and the activity index, a dynamic affiliation threshold is calculated for each cluster; The fusion vector is compared with the dynamic attribution threshold corresponding to the technology cluster with the highest similarity. If the similarity of the fusion vector is lower than the dynamic attribution threshold, the fusion vector is assigned to the candidate new cluster pool. If the similarity of the fusion vector is not lower than the dynamic attribution threshold, the fusion vector is added to the technology cluster with the highest similarity, and the center of the technology cluster is updated by using a moving average method.

5. The method according to claim 4, characterized in that, The method of updating the center of the technology cluster using a moving average is achieved by balancing the weights of the historical center of the technology cluster and the new patent vector through an adjustable decay factor.

6. The method according to claim 1, characterized in that, When the candidate new cluster pool meets the multi-dimensional triggering conditions, a large language model is invoked to perform multi-task in-depth analysis of the patents within the pool, generating an emerging technology early warning report, including: Monitor whether the growth in the number of patents in the candidate new cluster pool reaches a preset condition; Analyze whether the novelty of technical terms in the patents within the candidate new cluster pool exceeds a preset threshold; Whether the geographical distribution of patent applicants within the candidate new cluster pool meets a preset standard; When any one of the three dimensions of quantity growth, terminology novelty, and geographical distribution diffusion is met, a representative set of patents is selected from the candidate new cluster pool. Construct structured prompt templates to guide the large language model to perform three collaborative tasks: naming technology directions, predicting technology trends, and providing classification suggestions. The outputs of the three tasks are integrated into a structured early warning report on emerging technologies.

7. The method according to claim 6, characterized in that, The technology direction naming task is used to generate concise technology labels; the technology trend prediction task is used to output future technology development trends; and the classification suggestion task is used to output relevant patent classification numbers.

8. The method according to claim 5, characterized in that, The step of updating the clustering parameters required to calculate the dynamic threshold and optimizing the model parameters of the large language model based on the feedback from the target user includes: Record the type of action the target user takes in response to the emerging technology warning report, including confirmation, rejection, or modification. If the operation type is rejection, then the common features of the corresponding batch of patents are extracted and used to raise the threshold of the relevant triggering conditions. If the operation type is confirmation or modification, the technical information finally adopted by the target user will be used as a supervision tag. A training dataset is constructed using all feedback data, and a parameter-efficient fine-tuning method is used to optimize the large language model; Simultaneously, based on feedback information, the adjustment coefficient and the attenuation factor are dynamically adjusted.

9. The method according to claim 8, characterized in that, The initial values ​​of the adjustment coefficient and the attenuation factor are determined by an optimization algorithm during the system offline phase.

10. A self-evolving patent emerging technology discovery device, characterized in that, include: A receiving unit is used to receive newly input patent text, perform semantic embedding processing on the patent text, and generate a fusion vector; The processing unit is used to classify fusion vectors that are lower than the dynamic threshold into a candidate new cluster pool based on the similarity with each historical technology cluster center and a dynamic threshold calculated according to the cluster activity. The candidate new cluster pool is used to store fusion vectors that cannot be classified into any existing technology cluster. The generation unit is used to call a large language model to perform multi-task in-depth analysis of the patents in the pool when the candidate new cluster pool meets the multi-dimensional triggering conditions, and generate an emerging technology early warning report. The sending unit is used to push the emerging technology early warning report to the target user terminal and receive feedback from the target user; The adjustment unit is used to adjust the clustering parameters and the large language model analysis capability based on the target user feedback, and to update the clustering parameters required to calculate the dynamic threshold and the model parameters to optimize the large language model.

Images