Patent innovation degree evaluation method and device, and storage medium

Abstract

The application provides an evaluation method and device for the innovation degree of a patent and a storage medium. The evaluation method comprises the following steps: inputting a target patent and a patent data set, constraint semantic instructions and core semantic instructions into a pre-trained instruction control embedding model to obtain constraint semantic vectors and core semantic vectors corresponding to each patent respectively; determining a benchmark patent set within the same technical boundary as the target patent based on the constraint semantic vectors; determining a global comparison benchmark vector based on the core semantic vectors of each benchmark patent in the benchmark patent set; and evaluating the innovation degree of the target patent in the innovation investigation dimension based on the core semantic vector of the target patent and the global comparison benchmark vector.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and more specifically, to a method, apparatus, and storage medium for evaluating the degree of patent innovation. Background Technology

[0002] Patent innovation assessment is a core component of substantive patent examination and technical intelligence mining. With the explosive growth of global patent applications, traditional innovation examination methods that rely on manual expert searches or simple keyword matching are facing bottlenecks such as inefficiency and strong subjectivity.

[0003] Existing patent innovation identification schemes have significant limitations in practical applications. For example, evaluation methods based on external statistical indicators suffer from severe lag and uncertainty, making it impossible to conduct real-time assessments of the innovation of newly disclosed or pending patents. Identification schemes based on knowledge element extraction rely heavily on manually constructed domain dictionaries, rules, and knowledge graphs, resulting in complex extraction processes, high implementation costs, and poor generalization. General text embedding combined with anomaly detection analysis methods can only generate undifferentiated global semantic vectors, failing to achieve controllable semantic extraction for different innovation assessment dimensions. Evaluation results are easily affected by noise from irrelevant text, leading to poor interpretability.

[0004] Therefore, there is currently a lack of an evaluation scheme for patent innovativeness that offers controllable dimensions of creation, real-time performance, low cost, and high interpretability. Summary of the Invention

[0005] This application addresses the shortcomings of existing methods by proposing a method, apparatus, and storage medium for evaluating the degree of patent innovation, aiming to solve the problems of lag, high cost, and poor interpretability in existing patent innovation evaluation schemes.

[0006] Firstly, embodiments of this application provide a method for evaluating the degree of patent innovation, including: The target patent and patent dataset, along with constraint semantic instructions and core semantic instructions, are input into a pre-trained instruction control embedding model to obtain the constraint semantic vector and core semantic vector corresponding to each patent. The constraint semantic instructions are used to determine the technological boundaries, and the core semantic instructions are used to determine the dimensions of innovation assessment. Based on the constraint semantic vector, a set of benchmark patents that are within the same technical boundary as the target patent is determined; Based on the core semantic vectors of each benchmark patent in the benchmark patent set, a global comparison benchmark vector is determined. Based on the core semantic vector of the target patent and the global comparison benchmark vector, the degree of innovation of the target patent in the innovation examination dimension is evaluated. The patent dataset includes multiple patents from a specific time period, and these multiple patents belong to the same technical field as the target patent.

[0007] Secondly, embodiments of this application provide a device for evaluating the degree of patent innovation, comprising: The vector generation module is used to input the target patent and patent dataset, as well as the constraint semantic instructions and core semantic instructions, into the pre-trained instruction control embedding model to obtain the constraint semantic vector and core semantic vector corresponding to each patent. The constraint semantic instructions are used to determine the technological boundaries, and the core semantic instructions are used to determine the dimensions of innovation assessment. The determination module is used to determine a set of benchmark patents that are within the same technical boundary as the target patent based on the constraint semantic vector; and to determine a global comparison benchmark vector based on the core semantic vectors of each benchmark patent in the set of benchmark patents. The evaluation module is used to evaluate the degree of innovation of the target patent in the innovation examination dimension based on the core semantic vector of the target patent and the global comparison benchmark vector. The patent dataset includes multiple patents from a specific time period, and these multiple patents belong to the same technical field as the target patent.

[0008] Thirdly, embodiments of this application also disclose an electronic device, including a memory, a processor, and a computer program stored in the memory, characterized in that the processor executes the computer program to implement one or more of the evaluation methods described in the embodiments of the first aspect of this application.

[0009] Fourthly, embodiments of this application also disclose a computer-readable storage medium storing a computer program that, when executed by a processor, implements one or more of the evaluation methods described in the embodiments of the first aspect of this application.

[0010] Fifthly, embodiments of this application also disclose a computer program product, including a computer program that, when executed by a processor, implements one or more evaluation methods as described in the embodiments of the first aspect of this application.

[0011] The beneficial technical effects of the technical solutions provided in this application include: By using instruction-controlled embedding models to generate directional embedding vectors on specific semantic dimensions (i.e., core semantics and constraint semantics), while removing noise, a dynamic and controllable technical comparison benchmark is constructed. This can quantify the semantic shift on specific semantic dimensions, thereby improving the controllability, noise resistance, and interpretability of multi-dimensional technologies in patent innovation analysis. Therefore, it solves the problem of how to efficiently and automatically identify and evaluate patent innovation in an interpretable, high-precision, multi-dimensional, and semantically controllable manner.

[0012] Additional aspects and advantages of this application will be set forth in part in the description which follows, and will become apparent from the description or may be learned by practice of this application. Attached Figure Description

[0013] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 A flowchart illustrating a method for evaluating the degree of patent innovation provided in this application embodiment; Figure 2 A schematic diagram illustrating a method for evaluating the degree of patent innovation provided in an embodiment of this application; Figure 3 A schematic diagram of the structure of a patent innovation assessment device provided for an embodiment of this application; Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation The embodiments of this application are described below with reference to the accompanying drawings. It should be understood that the embodiments described below with reference to the accompanying drawings are exemplary descriptions for explaining the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions of the embodiments of this application.

[0014] Those skilled in the art will understand that, unless specifically stated otherwise, the terms "described" and "the" as used herein may also include plural forms. It should be further understood that the term "comprising" as used in this application's specification means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude implementations of other features, information, data, steps, operations, elements, components, and / or combinations thereof supported by this art. It should be understood that when we say an element is "connected" or "coupled" to another element, the element may be directly connected or coupled to the other element, or it may mean that the element and the other element are connected through an intermediate element. Furthermore, "connected" or "coupled" as used herein may include wireless connections or wireless coupling. The term "and / or" as used herein refers to at least one of the items defined by the term; for example, "A and / or B" may be implemented as "A," or as "B," or as "A and B."

[0015] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0016] With the explosive growth of global patent applications, traditional methods of examining patent innovation, which rely on manual expert searches or simple keyword matching, are facing bottlenecks such as inefficiency and high subjectivity. Currently, existing patent innovation identification technologies mainly include the following categories: The first category is innovation evaluation methods based on external statistical indicators. Examples include innovation analysis using indicators such as patent citation counts, patent family size, IPC classification span, and technology lifecycle. While these methods are relatively simple to implement, they suffer from significant lag. For instance, effective statistics on patent citations typically only become available a considerable time after patent publication or grant, making real-time identification of the innovation of newly published patents difficult. Furthermore, the reliability of the identification results is compromised due to the neglect of information inherent in the patent itself.

[0017] The second category is based on knowledge element extraction methods. These methods typically require extracting fine-grained knowledge units from patent texts, such as technical topics, technical terms, and functional structural relationships, and calculating innovativeness through knowledge recombination and semantic variation. While this method can improve semantic analysis capabilities to some extent, the knowledge extraction process is complex, the engineering implementation cost is high, and it relies heavily on domain dictionaries, rule templates, and knowledge graphs. Therefore, it not only has high computational costs but also struggles to generalize across domains.

[0018] The third category is based on text embedding and anomaly detection methods. For example, traditional text embedding models such as BERT and RoBERTa are used to generate patent vectors, which are then combined with Local Outlier Factor (LOF), cluster analysis, or anomaly detection algorithms to assess the degree of patent innovation. This type of method can reduce reliance on manual rules to some extent, but traditional text embedding typically only generates an overall representation in a unified semantic space, making it difficult to perform controllable analysis on specific innovation dimensions. Therefore, it suffers from problems such as semantic uninterpretability, uncontrollable innovation dimensions, and significant noise interference.

[0019] Specifically, traditional text embedding models (such as BERT and RoBERTa) lack instruction-following capabilities and can only generate "global / overall semantic embeddings" of patent texts. In real-world scenarios, the semantic similarity of a patent based on the "technical solution" dimension and the "application scenario" dimension is often very small, and conventional embedding cannot extract features with different focuses according to specific analysis needs.

[0020] At the same time, overall embedding is prone to introducing a large amount of background noise information that is not related to innovation (such as industry-standard terms and legal jargon). Furthermore, due to the uncontrollable vector space, the output "outlier" or similarity calculation results are difficult to provide examiners or R&D personnel with specific, multi-dimensional, and technically reasonable explanations.

[0021] In recent years, text embedding models with instruction-following capabilities have gradually emerged. Models such as Qwen3-Embedding and BGE can receive prompts or instructions during the embedding process, thereby generating vector representations with different semantic emphases based on different analysis needs. However, a patent innovation identification method based on instruction embedding mechanisms that can achieve controllable semantic analysis is currently lacking.

[0022] Therefore, there is an urgent need to propose a patent innovation identification (which can also be described as judgment, determination, evaluation, etc.) method that can dynamically control the embedded semantic content according to different innovation analysis objectives and has strong interpretability and controllability.

[0023] To address the aforementioned shortcomings, this application proposes a patent innovation identification method based on instruction embedding. It utilizes an instruction-controlled embedding model to generate directional vectors along specific semantic dimensions (core semantics and constraint semantics). This removes noise while constructing a dynamic and controllable technical comparison benchmark, thereby achieving highly interpretable quantification of innovation.

[0024] The technical solution of this application and how it solves the above-mentioned technical problems are described in detail below with specific embodiments. It should be noted that the following embodiments can be referenced, borrowed, or combined with each other, and the same terms, similar features, and similar implementation steps in different embodiments will not be described again.

[0025] The following is a detailed description of the patent innovation assessment method and apparatus proposed in the embodiments of this application, with reference to the accompanying drawings.

[0026] In some embodiments, a method for assessing the degree of patent innovation is provided. For example... Figure 1 As shown, the evaluation method includes: S10. Input the target patent and patent dataset, as well as the constraint semantic instructions and core semantic instructions, into the pre-trained instruction control embedding model to obtain the constraint semantic vector and core semantic vector corresponding to each patent.

[0027] Optionally, the constraint semantic instructions are used to determine the technology boundaries, and the core semantic instructions are used to determine the innovation assessment dimensions. Optionally, the patent dataset includes multiple patents within a specific time period, and these multiple patents belong to the same technical field as the target patent.

[0028] It should be understood that the target patent in this application embodiment is a patent whose inventiveness is to be examined or analyzed. The patent dataset is a collection of comparative patents used to analyze the inventiveness of the target patent.

[0029] In some embodiments, two differentiated semantic instructions are preset for the target patent: a constraint semantic instruction is used to define the boundaries of the unified technology application, and a core semantic instruction is used to lock in the dimensions of innovation assessment.

[0030] Optionally, core semantic instructions can be any basic dimensional instruction configured during model training, but are not limited to this. Constraint semantic instructions can be derived from the basic dimensional instructions configured during model training, but are not limited to this.

[0031] In some embodiments, a single patent and its corresponding semantic instructions are input into the instruction control embedding model. Optionally, the target patent and all prior art patents within a specified publication period are paired with constraint semantic instructions and input into the instruction control embedding model, outputting constraint semantic embedding vectors for each patent; then, all the aforementioned patents are paired with core semantic instructions and input into the instruction control embedding model, outputting core semantic embedding vectors for each patent.

[0032] In some embodiments, before step S10, the step may further include: training an instruction control embedding model.

[0033] Optionally, the instruction control embedding model can be trained through the following process: S01. Input the training dataset and semantic instructions of at least two dimensions into the initial instruction embedding model. The model performs forward inference and outputs the semantic instruction embedding vector corresponding to each training sample in the training dataset. Optionally, each training sample in the training dataset includes at least: the title and abstract of the patent. Optionally, at least two dimensions of semantic instructions can be understood as basic dimension instructions, such as structural feature dimension instructions and application scenario dimension instructions, but not limited to these.

[0034] Optionally, text data containing patent titles and abstracts is collected and aggregated to form a training dataset, and at least two dimensions of semantic instructions are configured. Then, a single patent sample in the training dataset is paired with a single-dimensional semantic instruction and input into an initial instruction embedding model. The model only performs forward inference and outputs the semantic instruction embedding vector corresponding to each patent sample. This embedding vector is an intermediate result of the model's computation.

[0035] Optionally, for each training sample, a corresponding semantic instruction embedding vector can be generated for each dimension of the input semantic instruction. That is, each training sample can correspond to at least two semantic instruction embedding vectors.

[0036] S02. Based on the semantic instruction embedding vectors corresponding to each training sample, determine the third similarity between training samples in the training dataset.

[0037] Optionally, cosine similarity between samples can be calculated based on the first semantic instruction embedding vector of all training samples in the training dataset. Here, the first semantic instruction embedding vector is generated by the model based on semantic instructions of a certain dimension.

[0038] Optionally, a tuple can be randomly selected from the training dataset (e.g., every 3 training samples form a tuple), and the similarity of the first semantic instruction embedding vector of that tuple under a certain dimension of semantic instruction can be calculated. This method yields the cosine similarity among all training samples in the training dataset.

[0039] S03. Based on the third similarity, construct a triplet training sample set. Optionally, each triplet training sample in the triplet training sample set includes: anchor sample, positive sample, and negative sample.

[0040] Optionally, similar samples (positive examples) and dissimilar samples (negative examples) are initially distinguished based on similarity values. Then, the judgment bias is corrected by manual verification. Based on the similarity results after verification, any patent in the tuple is used as the anchor sample. Positive examples (those with the highest similarity to the anchor sample) and negative examples (those with the lowest similarity to the anchor sample) are paired to form triplet training samples. Then, a triplet training sample set consisting of anchor samples, positive examples, and negative examples is constructed.

[0041] S04. Perform multiple rounds of iterative training based on a deep learning framework and an adaptive gradient optimizer. During the training process, input the triplet sample set into the initial instruction embedding model and use the joint loss function to backpropagate and update the parameters of the initial instruction embedding model to obtain the instruction control embedding model.

[0042] Optionally, the joint loss function is formed by a weighted fusion of the first loss function and the second loss function. For example, the joint loss function (Loss) can be determined by the following formula:

[0043] in, To adjust the first loss function With the second loss function The balance coefficient of the loss weight takes a value between 0 and 1.

[0044] Optionally, the first loss function represents the instruction-specific contrastive loss, which is used to reduce the vector distance between the anchor sample and the positive sample and increase the vector distance between the anchor sample and the negative sample in the vector space corresponding to a single instruction, so that the model learns the ability to extract subdivided semantics in accordance with the question instruction.

[0045] For example, the first loss function L1 can be determined by the following formula:

[0046] Where τ is a temperature hyperparameter used to control the gradient propagation scale for similarity calculation. For the first This patent contains embedded instructions The instruction condition embedding vector generated under guidance. For the first The patent in the directive The instruction condition embedding vector of the corresponding similar patent (positive example). For the first The patent in the directive The instruction condition embedding vector of the corresponding dissimilar patent (negative example).

[0047] Optionally, the second loss function represents the global embedding regularization contrastive loss, which is used to maintain the relationship between the vector distances of anchor samples, positive samples, and negative samples in a general vector space without instructions, so that the model avoids overfitting and retains its generalization ability to general semantics.

[0048] For example, the second loss function L2 can be determined by the following formula:

[0049] Where τ is a temperature hyperparameter used to control the gradient propagation scale for similarity calculation. For the first The patent describes the original global embedding vector in the absence of instructions. For the first The patent in the directive The instruction condition embedding vector of the corresponding similar patent (positive example). For the first The patent in the directive The instruction condition embedding vector of the corresponding dissimilar patent (negative example).

[0050] Optionally, a temperature hyperparameter configured during training can be used to adjust the gradient scaling scale for similarity calculation between semantic instruction embedding vectors, in order to balance the convergence speed of the loss function with the discrimination accuracy.

[0051] Optionally, the loss function includes the first loss function and the second loss function.

[0052] It should be noted that the deep learning framework in this application embodiment can be a deep learning computing framework used to build a model training pipeline, realizing batch loading of samples, vector tensor operations, loss calculation, and parameter iterative updates. For example, but not limited to, the PyTorch framework.

[0053] It should also be noted that the adaptive gradient optimizer in the embodiments of this application can be an adaptive gradient optimizer with independent weight decay, used to adaptively adjust model parameters and suppress model overfitting based on the loss gradient. For example, but not limited to, the AdamW optimizer.

[0054] S11. Based on the constraint semantic vector, determine the set of benchmark patents that are within the same technical boundary as the target patent.

[0055] In some embodiments, step S11 may specifically include: determining the benchmark patent set based on the first similarity between the constraint semantic vector corresponding to the target patent and the constraint semantic vector corresponding to each patent in the patent dataset.

[0056] In some embodiments, the benchmark patent set can be determined through the following process: Calculate the first similarity between the constraint semantic vector corresponding to the target patent and the constraint semantic vector corresponding to each patent in the patent dataset; After sorting the patents by first similarity from high to low, the top k patents with the highest first similarity are selected to form the benchmark patent set, where k is an integer greater than or equal to 2.

[0057] For example, assuming the patent dataset includes 1000 comparative patents, the similarity between the constraint semantic vector corresponding to the target patent and the constraint semantic vector corresponding to each of these 1000 comparative patents is calculated, resulting in 1000 similarity scores. These 1000 similarity scores are then sorted from highest to lowest, and the top 5 comparative documents with the highest similarity scores are selected as benchmark patents. These 5 comparative documents constitute the benchmark patent set.

[0058] It should be noted that the numbers in this example are merely examples to illustrate the solution and do not constitute any limitation on the solution of this application.

[0059] S12. Based on the core semantic vectors of each benchmark patent in the benchmark patent set, determine the global comparison benchmark vector.

[0060] In some embodiments, step S12 may specifically include: determining the global comparison benchmark vector based on the first similarity and core semantic vector corresponding to each benchmark patent.

[0061] In some embodiments, the global comparison benchmark vector can be determined through the following process: The first similarity corresponding to each of the benchmark patents is normalized to obtain the normalized weighted coefficients corresponding to each of the benchmark patents. Based on the normalized weighting coefficients corresponding to each of the benchmark patents, the corresponding core semantic vectors are weighted and fused to obtain a fused vector; The fusion vector is normalized using Euclidean norm to obtain the global comparison benchmark vector.

[0062] Optionally, the global comparison benchmark vector can be obtained using the following formula:

[0063] in, This is done by normalizing the Euclidean norm. For the j-th benchmark patent, the normalized weighting coefficient is... Let be the core semantic embedding vector corresponding to the j-th benchmark patent.

[0064] Optionally, the first similarity scores corresponding to each benchmark patent can be calculated. The normalization process yields the normalized weighted coefficients for each benchmark patent. Specifically, the implementation of Softmax normalization for the similarity scores of benchmark patents can include: performing an exponential transformation on each similarity score to make all values ​​positive and increase the difference in similarity scores; then summing all the transformed values ​​and dividing each transformed value by the summation to obtain a normalized weighted coefficient with a value between 0 and 1, and a sum of all weights equal to 1. Benchmark patents with higher similarity scores are assigned larger weights.

[0065] S13. Based on the core semantic vector of the target patent and the global comparison benchmark vector, evaluate the degree of innovation of the target patent in the innovation examination dimension.

[0066] In some embodiments, step S13 may specifically include: Calculate the second similarity between the core semantic vector corresponding to the target patent and the global comparison benchmark vector; Based on the second similarity and the preset threshold, the degree of innovation of the target patent in the innovation examination dimension is determined.

[0067] In some embodiments, the global comparison benchmark vector is used to characterize the overall technological distribution of benchmark patents in the innovation examination dimension within the same technological boundary.

[0068] Optionally, the innovativeness of the target patent in the innovative examination dimension determined by the core semantic instructions can be judged (or evaluated) by the similarity between the core semantic vector corresponding to the target patent and the global comparison benchmark vector.

[0069] In some alternative embodiments, the preset threshold may include a first threshold and a second threshold, wherein the first threshold is greater than the second threshold.

[0070] Optionally, when the second similarity is greater than or equal to the first threshold, the core dimension technical features of the target patent are determined to be highly overlapping with the prior art, and the degree of innovation (or innovation) is low; when the second similarity is less than or equal to the second threshold, the core dimension technical features of the target patent are determined to deviate from the conventional distribution of the prior art, and the degree of innovation is high.

[0071] Optionally, the first threshold can be 0.9 and the second threshold can be 0.45, but are not limited to these.

[0072] It should be noted that the "similarity" involved in the embodiments of this application can be "cosine similarity", but is not limited to this.

[0073] In some embodiments, combined with Figure 2This application provides a detailed description of a method for evaluating the degree of patent innovation. This evaluation method may include the following steps: S1. Construct a patent instruction embedding dataset and perform comparative learning domain fine-tuning on the model.

[0074] Classic text embedding models cannot provide tailored embeddings to specific patent texts based on analytical needs. Recent open-source embedding models (such as Qwen3-Embedding and BGE) can receive prompt parameters during the embedding process to guide the generation of embedding vectors, exhibiting a certain degree of instruction compliance. To adapt the model to the specialized expressions within the patent domain, this step involves fine-tuning and optimizing the model through the following process: 1. Data Collection: Collect patent text data for fine-tuning the model, including at least the patent title and abstract.

[0075] 2. Analysis Instruction Setting: Set analysis instructions for different semantic dimensions based on the patent text. For example, set "What is the application scenario of this patent?" for application scenarios, and "What is the core architecture of the technical solution of this patent?" for technical structures.

[0076] 3. Sample pre-extraction: In the patent text data, the overall similarity is first calculated based on the untuned embedding model. In the pairwise patent combinations with high similarity, the three patents with the highest similarity are extracted to construct tuples.

[0077] 4. Positive and negative example labeling: Introduce manual or large model, and according to the specific analysis instructions set in step 2, perform targeted labeling on the above-mentioned high similarity patent tuples (e.g., in the dimension of "application scenario", patent A is closer to patent C than patent B, then we get anchor sample (Anchor) patent C, positive sample (Positive) patent A, and negative sample (Negative) patent B).

[0078] 5. Contrastive Learning Fine-tuning: In the deep learning framework PyTorch, a custom contrastive loss function is used to fine-tune the initial model through forward and backward propagation to optimize the model parameters.

[0079] The joint loss function designed in this embodiment is... and It is composed of weighted combinations, and the specific formula is as follows: (1) (2) (3) The symbols in the above formula have the following meanings: : No. This patent contains embedded instructions The instruction condition embedding vector generated under guidance.

[0080] : No. The patent in the directive The instruction condition embedding vector of the corresponding similar patent (positive example).

[0081] : No. The patent in the directive The instruction condition embedding vector of the corresponding dissimilar patent (negative example).

[0082] : No. The patent describes the original global embedding vector in the absence of instructions.

[0083] τ: Temperature hyperparameter controlling the gradient propagation scale.

[0084] :adjust and The balance coefficient of the loss weight takes a value between 0 and 1.

[0085] Explanation of the above formula: The naive InfoNCE loss introduced after embedding instructions aims to narrow the distance between positive examples and widen the distance between negative examples under the same instructions, thereby giving the model the ability to follow instructions.

[0086] This application proposes an innovative assumption based on its embodiments: the partial order similarity relationship between instruction embeddings also applies to the original embedding of anchor patents. Since instruction embeddings represent the fine-grained semantics of a patent in a specified dimension, while the original embedding represents the overall semantics, the overall semantics should encompass some information of that specified dimension, making its similarity calculation result approximate that of the instruction embedding. (Introduction) It can play a regularization role, preventing the model from overfitting or losing its ability to generalize the semantics of the basic patent during the fine-tuning process.

[0087] S2. Use a fine-tuning model to perform targeted control embedding of the core semantics and constraint semantics of the patent to be analyzed.

[0088] The patent data to be analyzed can be a collection of patents in a specific technical field during a specific period (e.g., graphene material patents publicly granted from January to June 2020) to ensure that the patents have basic architectural comparability.

[0089] Based on this, embodiments of this application can control the embedding model through two different instructions to extract two key semantic vectors respectively: 1. Core semantic embedding: Used to represent the specific technical innovation points of the patent that are of primary concern in innovation analysis. Core semantic instructions are used to guide the model extraction.

[0090] For example, if the focus is on innovations in the structural dimension of graphene, instructions can be used. =“What is the specific structure of the graphene material in the following patent?”, generate vector .

[0091] 2. Constraint Semantic Embedding: Used to characterize the semantic boundaries of selected innovative comparative patents. The innovativeness of a patent is revealed in comparison with similar technologies, and constraint semantic instructions are used to describe which patents are substantially comparable.

[0092] For example, if the focus is on composite graphene materials, instructions can be used. =“How is graphene used in composite materials in the following patents?”, generating vector .

[0093] In some embodiments, the distance between constraint semantics in the vector space is inversely proportional to the comparability between patents.

[0094] S3. Obtain a dynamic innovative benchmark document set based on constraint semantic retrieval.

[0095] To assess the innovativeness of a target patent, a dynamic pool of prior art needs to be tailored to that target patent. Specifically, the prior art pool can be determined in the following ways: 1. Traverse the patent dataset to be analyzed, based on the constraint semantic embedding vector generated in step S2. Calculate the cosine similarity between the target patent and all other patents in the dataset.

[0096] 2. Sort the samples from highest to lowest similarity and select the samples with the highest similarity. These patents serve as a dynamic benchmark document set for comparing the innovativeness of the target patent.

[0097] The following require instructions: constants The value of is determined by the specific analytical requirements, and its magnitude is negatively correlated with the consistency (technology concentration) within the patent dataset to be analyzed. The more concentrated the technology in the dataset, the better. The smaller the value, the better.

[0098] S4. Construct a global contrastive semantic representation vector.

[0099] To avoid bias (or errors) caused by a single comparison file, this step will The semantics of the benchmark patents are weighted and fused to form a virtual overall status quo vector representing the current technological background. This vector can be obtained through the following process: 1. Extract the above The semantic similarity of the benchmark patents is normalized using the Softmax function to calculate the normalized weight coefficient for each benchmark patent.

[0100] 2. Using this weighting coefficient, for this The weighted average of the core semantic embedding vectors corresponding to the benchmark patents is calculated to obtain the global core semantic weighted average vector.

[0101] 3. Perform the following steps on the weighted mean vector: After normalization, a comparative semantic embedding vector corresponding to the target patent is finally generated. This vector accurately represents the overall state of the prior art for a certain core technical feature (reflected by the core semantics) under a specified technical boundary (controlled by constraint semantics).

[0102] It should be noted that, Normalization is also known as Euclidean norm normalization or normalization. Its core function is to scale a set of numerical values ​​(or a vector) of arbitrary length proportionally so that the sum of the squares of all elements equals 1 (i.e., the geometric length / magnitude of the vector becomes 1), while keeping the original vector's direction completely unchanged.

[0103] It should also be noted that the L2 here is different from the L2 in step S1 above. The L2 in step S1 is the loss function, while the L2 here is a normalization method.

[0104] S5. Calculate the innovation score to achieve interpretable quantitative assessment.

[0105] 1. Calculate the core semantic embedding of the target patent itself. The cosine similarity between its corresponding contrast semantic embedding vectors.

[0106] 2. This similarity is inversely proportional to the degree of difference between the target patent and the core technical features in related fields: (1) The higher the similarity, the smaller the difference between the core technical content (such as specific structure) of the target patent in the innovation analysis and the overall status of the existing technology, indicating that the innovation in this dimension is low; (2) The lower the similarity, the more it indicates that the core technical features of the target patent deviate significantly from the weighted distribution of the prior art, indicating that the innovation in this dimension is high.

[0107] 3. Since the instructions are user-controllable (such as issuing different core instructions such as process steps and material ratios), by viewing the similarity scores under different instructions, it is possible to clearly and interpretably output the specific dimension in which the target patent is innovative.

[0108] Therefore, compared with existing traditional solutions, the solution of this application embodiment can bring the following beneficial effects to patent analysis tools or examination systems: 1. High efficiency and low cost, eliminating cumbersome information extraction pipelines: This invention eliminates the need for time-consuming, costly, and poorly generalized fine-grained knowledge element (such as entity and relation) extraction and explicit knowledge network construction. It employs instruction embedding to directly capture specific-dimensional semantic features from the original text, significantly reducing the data preprocessing flow and enabling rapid batch processing of large-scale patent databases.

[0109] 2. Multi-dimensional and highly versatile controllable analysis experience: Users (reviewers or enterprise R&D personnel) do not need to change the underlying model. They can control the system to focus on different technical evaluation emphases (such as choosing to focus on formulation innovation or equipment structure innovation) simply by designing and adjusting the Prompt descriptions of core semantic instructions and constraint semantic instructions. The system demonstrates strong customization and scenario generalization capabilities.

[0110] 3. Extremely high interpretability and noise resistance: Traditional text vector outlier detection methods are black boxes and easily affected by noise such as legal jargon within patent texts. The solution in this application's embodiment constrains semantic locking to compare broad category boundaries, examines detailed items through core semantic locking technology, and finally quantifies innovativeness through inverse similarity logic. This not only accurately filters background noise but also provides highly instructive, multi-dimensional interpretable reports, such as those showing 85% uniqueness in the structural dimension of a patent, yet exhibiting a high degree of overlap with existing technologies in application scenarios.

[0111] Furthermore, existing knowledge element extraction methods rely on external information extraction models or finely defined rules to extract fine-grained explicit knowledge units (such as entities and phrases), making the system costly for processing complex long texts and cross-domain transfer. In contrast, the solution in this application directly employs instruction embedding technology, endowing the vector model with instruction-following capabilities to directly filter out the required deep implicit semantic information from the patent text. This not only avoids complex text structure extraction processes but also, because the embedded vectors contain richer contextual semantic details (rather than simple technical phrases and symbols), results in stronger generalization and higher accuracy in innovation assessment.

[0112] Existing methods based on conventional text embedding and outlier detection, using conventional language models such as RoBERTa, generate semantic embeddings that are "all-encompassing" global vectors. These vectors cannot distinguish the semantic emphasis of a patent in terms of "technical solution" or "application scenario," are easily mixed with noise such as legal terminology and industry background information, and the final calculated "local outlier score" lacks technical interpretability. The solution in this application, however, utilizes instruction embedding to simultaneously introduce "core semantic instructions" and "constraint semantic instructions." The constraint semantic instructions dynamically control the semantic space range of the comparison benchmark (defining the comparable technology pool), while the core semantic instructions control the specific dimensions of the examined features (removing redundant noise). The comparison benchmark vector integrated through this dual-instruction control mechanism can directly quantify the semantic shift in specific dimensions, thereby significantly improving the controllability, noise resistance, and multi-dimensional technical interpretability of innovation analysis.

[0113] Therefore, the solution in this application addresses the problem of how to efficiently and automatically identify and evaluate patent innovativeness in an interpretable, high-precision, multi-dimensional, and semantically controllable manner. It can be directly applied to: 1. National or institutional patent novelty / inventory examination assistance system: automatically filters comparison documents and quantifies the degree of innovation.

[0114] 2. Enterprise Technological Innovation Frontier Identification and Patent Navigation Project: Discovering breakthrough technologies and technological gaps in specific fields.

[0115] 3. Tools for evaluating the commercialization of scientific and technological achievements and measuring the value of intellectual property: providing quantitative proof of the innovative value of technological assets.

[0116] The following describes in detail the patent innovation assessment scheme of this application embodiment, using application examples of the scheme.

[0117] This embodiment takes the evaluation of a batch of patents to be analyzed in the field of "graphene composite materials" as an example to describe in detail the method of innovativeness identification (or evaluation) through the embodiment of this application.

[0118] The method may include: Step 1: Constructing the dataset and fine-tuning the model (corresponding to step S1 in the above embodiment).

[0119] 1. Data collection: Retrieve and download the titles and abstracts of 50,000 patents in the field of graphene technology from the patent database. 2. Set analysis commands: Set commands for two dimensions for this domain: Structural feature dimension instructions "What is the specific structure of the graphene material in the following patent?" Application scenario dimension instructions "What are the application scenarios for this patent?"

[0120] 3. Sample extraction and labeling: The basic model BGE-M3 was selected as the initial instruction embedding model. Randomly select a tuple (e.g., patent X, patent Y, patent Z). In the instruction... The original similarity is calculated based on structural features. After verification by a large model or manual assistance, it is found that: structurally, patent X is very similar to patent Y, but is fundamentally different from patent Z. This leads to the construction of triples: anchor samples For patent X, positive example sample For patent Y, negative sample Patent Z. 4. Model Fine-tuning: Within the PyTorch framework, feed the above samples into the model. Utilize the joint loss function... (Pick The model is trained iteratively for 3 epochs by backpropagation through the AdamW optimizer, enabling it to acquire strong instruction compliance with graphene patent texts.

[0121] Step 2: Targeted control embedding of the patent to be analyzed (corresponding to step S2 in the above embodiments).

[0122] There is a newly submitted patent application, patent A, whose abstract reads: "A composite film consisting of alternating layers of graphene and titanium dioxide nanoparticles, primarily used as an electrode material for high-frequency energy storage capacitors...". To evaluate the innovativeness of patent A in the "material structure" dimension, and to constrain its "comparison boundary" within the "composite material application" dimension, the following dual instructions are set: 1. Constraining semantic instructions (defining technical boundaries): =“How is graphene used in composite materials in the following patents?”; 2. Core semantic instructions (defining the dimensions of examination): = "What is the specific structure of the graphene material in the following patent?"

[0123] Input the target patent A, along with 1000 other graphene-related patents from specific time periods (e.g., published from January to June 2024) included in the patent dataset, into the fine-tuned model: Obtain the constraint semantic embedding vector of target patent A . Obtain the core semantic embedding vector of target patent A . Using the same method, the constraint semantic vectors and core semantic vectors of the other 1,000 comparative patents were obtained.

[0124] Step 3: Retrieve the dynamic innovation comparison benchmark document set (corresponding to step S3 in the above embodiment).

[0125] 1. Using the constraint semantic vector of target patent A The cosine similarity is calculated for each of the constraint semantic vectors of the other 1,000 comparative patents. 2. Sort by similarity from highest to lowest, and set a constant. . 3. Select the 5 patents with the highest similarity (e.g., denoted as patents). arrive The abstracts of these five patents all heavily involve schemes for "using graphene in composite materials," forming a dynamic benchmark pool for comparing innovation against target patent A. Step 4: Construct a global contrastive semantic representation vector (corresponding to step S4 in the above embodiment).

[0126] 1. Extract the constraint semantic similarity scores between these 5 comparative patents and target patent A, assuming the following: . 2. Perform the following steps on the score array: Normalization, calculate the corresponding weighting coefficients. . 3. Extract the core semantic embedding vectors generated in step two for these five comparative patents (i.e., the vectors generated by each of the five comparative patents under the "Structure" instruction). ). 4. Perform weighted summation and then... Normalization is performed to calculate the contrastive semantic embedding vector.

[0127] This vector accurately represents the current state of existing technologies regarding "material structure" within the technological boundary of "graphene composite materials".

[0128] Step 5: Calculate the innovation score and analyze the results (corresponding to step S5 in the above embodiment).

[0129] 1. Calculate the core semantic embedding of target patent A Contrastive semantic embeddings with newly generated Cosine similarity between them. 2. Scenario Result Deduction: (1) Scenario 1 (Low Innovation): If the calculated cosine similarity is 0.91 (high similarity), it means that the "alternating assembly structure of layered graphene and nanoparticles" proposed in target patent A has been widely adopted by similar technologies under the current boundary of composite materials, with minimal differences. The system outputs that its innovation in the structural dimension is low. (2) Scenario 2 (High Innovation): If the calculated cosine similarity is 0.42 (low similarity), it means that although they all belong to the boundary of composite materials (constraint semantics are similar), the specific "alternating assembly three-dimensional structural features" of target patent A deviates from the conventional distribution of existing technologies, and the difference is significant. The system outputs that it has high innovation in the structural dimension. Through this embodiment, the system not only provides a quantifiable innovation score, but also clearly explains that the score is calculated under the specific instruction dimension of "structure", thus realizing the automatic identification of controllable and explainable patent innovation (or degree of innovation).

[0130] Based on the same inventive concept, embodiments of this application provide a device for evaluating the degree of patent innovation. The following, in conjunction with... Figure 3 The patent innovation assessment device provided in this application embodiment is described in detail.

[0131] like Figure 3 As shown, the evaluation device 20 includes a vector generation module 21, a determination module 22, and an evaluation module 23. Among them, The vector generation module 21 is used to input the target patent and patent dataset, as well as the constraint semantic instructions and core semantic instructions into the pre-trained instruction control embedding model, to obtain the constraint semantic vector and core semantic vector corresponding to each patent respectively. The determination module 22 is used to determine a set of benchmark patents that are within the same technical boundary as the target patent based on the constraint semantic vector; and to determine a global comparison benchmark vector based on the core semantic vector of each benchmark patent in the set of benchmark patents. Evaluation module 23 is used to evaluate the degree of innovation of the target patent in the innovation examination dimension based on the core semantic vector of the target patent and the global comparison benchmark vector; The constraint semantic instructions are used to determine the technological boundaries, and the core semantic instructions are used to determine the dimensions of innovation assessment. The patent dataset includes multiple patents from a specific time period, and these multiple patents belong to the same technical field as the target patent.

[0132] In some optional embodiments, when determining the set of benchmark patents within the same technical boundary as the target patent based on the constraint semantic vector, the determining module 22 is specifically used for: The benchmark patent set is determined based on the first similarity between the constraint semantic vector corresponding to the target patent and the constraint semantic vector corresponding to each patent in the patent dataset.

[0133] In some optional embodiments, the determining module 22, when determining the benchmark patent set, is specifically used for: Calculate the first similarity between the constraint semantic vector corresponding to the target patent and the constraint semantic vector corresponding to each patent in the patent dataset; After sorting the patents by first similarity from high to low, the top k patents with the highest first similarity are selected to form the benchmark patent set, where k is an integer greater than or equal to 2.

[0134] In some optional embodiments, when determining the global comparison benchmark vector based on the core semantic vectors of each benchmark patent in the benchmark patent set, the determining module 22 is specifically used for: The global comparison benchmark vector is determined based on the first similarity and core semantic vector corresponding to each benchmark patent.

[0135] In some optional embodiments, the determining module 22, when determining the global comparison reference vector, is specifically used for: The first similarity corresponding to each of the benchmark patents is normalized to obtain the normalized weighted coefficients corresponding to each of the benchmark patents. Based on the normalized weighting coefficients corresponding to each of the benchmark patents, the corresponding core semantic vectors are weighted and fused to obtain a fused vector; The fusion vector is normalized using Euclidean norm to obtain the global comparison benchmark vector.

[0136] In some optional embodiments, when evaluating the degree of innovation of the target patent in a specified dimension based on the core semantic vector of the target patent and the global comparison benchmark vector, the evaluation module 23 is specifically used for: Calculate the second similarity between the core semantic vector corresponding to the target patent and the global comparison benchmark vector; Based on the second similarity and the preset threshold, the degree of innovation of the target patent in the innovation examination dimension is determined.

[0137] In some alternative embodiments, it further includes: a training module for training the instruction control embedding model through the following process: The training dataset and semantic instructions of at least two dimensions are input into the initial instruction embedding model. The model performs forward inference and outputs the semantic instruction embedding vector corresponding to each training sample in the training dataset. Each training sample in the training dataset includes at least: the title and abstract of the patent. Based on the semantic instruction embedding vectors corresponding to each training sample, the third similarity between training samples in the training dataset is determined. Based on the third similarity, a triplet training sample set is constructed; wherein, each triplet training sample in the triplet training sample set includes: anchor sample, positive sample, and negative sample; The training is performed in multiple rounds based on a deep learning framework and an adaptive gradient optimizer. During the training process, the triplet sample set is input into the initial instruction embedding model, and the parameters of the initial instruction embedding model are updated by backpropagation using a joint loss function to obtain the instruction control embedding model. The joint loss function is formed by a weighted fusion of the first loss function and the second loss function; The first loss function represents the instruction-specific contrastive loss, which is used to reduce the vector distance between the anchor sample and the positive sample and increase the vector distance between the anchor sample and the negative sample in the vector space corresponding to a single instruction, so that the model learns the ability to extract subdivided semantics in accordance with the question instruction. The second loss function represents the global embedding regularization contrastive loss, which is used to maintain the distance between anchor samples, positive samples, and negative samples in a general vector space without instructions, so that the model avoids overfitting and retains the ability to generalize to general semantics.

[0138] In some alternative embodiments, the temperature hyperparameter configured during training is used to adjust the gradient scaling scale for similarity calculation between semantic instruction embedding vectors in order to balance the convergence speed of the loss function with the discrimination accuracy. The loss function includes the first loss function and the second loss function.

[0139] In some alternative embodiments, the global comparison benchmark vector is used to characterize the overall technological distribution of benchmark patents in the innovation dimension within the same technological boundary.

[0140] It should be understood that the modules in the above embodiments may include processors. A processor may include a CPU (Central Processing Unit), a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field-Programmable Gate Array), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this application. A processor may also be a combination that implements computational functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, etc.

[0141] The evaluation apparatus provided in this application embodiment can achieve the above-mentioned... Figure 1 and Figure 2 The various processes implemented in the evaluation method embodiments shown are not described again here to avoid repetition.

[0142] Based on the same principles as the methods shown in the embodiments of this application, the embodiments of this application also provide an electronic device, which may include a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps and corresponding contents of the aforementioned evaluation method embodiments.

[0143] In an alternative embodiment, an electronic device, such as Figure 4 Show. Figure 4 The illustrated electronic device 30 includes a processor 31 and a memory 32. The processor 31 is communicatively connected to the memory 32, for example, via a bus 33. Optionally, the electronic device 30 may further include a transceiver 34, which can be used for data interaction between the electronic device and other electronic devices, such as sending and / or receiving data. It should be noted that in practical applications, the transceiver 34 is not limited to one type, and the structure of the electronic device 30 does not constitute a limitation on the embodiments of this application.

[0144] Processor 31 may be a CPU (Central Processing Unit), a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field-Programmable Gate Array), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof, comprising the chip or trigger pulse described in any of the above embodiments. It may implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this application. Processor 31 may also be a combination that implements computational functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, etc.

[0145] Bus 33 may include a pathway for transmitting information between the aforementioned components. Bus 33 may be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc. Bus 33 can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 4 It is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0146] The memory 32 may be a ROM (Read-Only Memory) or other type of static storage device capable of storing static information and instructions, RAM (Random Access Memory) or other type of dynamic storage device capable of storing information and instructions, or it may be an EEPROM (Electrically Erasable Programmable Read Only Memory), a CD-ROM (Compact Disc Read-Only Memory) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto.

[0147] The memory 32 is used to store computer programs that execute the embodiments of this application, and the execution is controlled by the processor 31. The processor 31 is used to execute the computer programs stored in the memory 32 to implement the steps shown in the foregoing method embodiments.

[0148] This application provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it can implement the steps and corresponding content of the aforementioned method embodiments.

[0149] This application also provides a computer program product, including a computer program that, when executed by a processor, can implement the steps and corresponding content of the aforementioned method embodiments.

[0150] In the description of this application, the terms "center," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicate directions or positional relationships based on the exemplary directions or positional relationships shown in the accompanying drawings. They are used to facilitate the description or simplification of the embodiments of this application and are not intended to indicate or imply that the device or component referred to must have a specific orientation or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application.

[0151] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, unless otherwise stated, "a plurality of" means two or more.

[0152] In the description of this application, it should be noted that, unless otherwise expressly specified and limited, the terms "installation," "connection," and "joining" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0153] In the description of this specification, specific features, structures, materials, or characteristics may be combined in any suitable manner in one or more embodiments or examples.

[0154] The above description is only a partial implementation of this application. It should be noted that for those skilled in the art, other similar implementation methods based on the technical concept of this application, without departing from the technical concept of this application, also fall within the protection scope of the embodiments of this application.

Claims

1. A method for evaluating the degree of patent innovation, characterized in that, include: The target patent and patent dataset, along with constraint semantic instructions and core semantic instructions, are input into a pre-trained instruction control embedding model to obtain the constraint semantic vector and core semantic vector corresponding to each patent. The constraint semantic instructions are used to determine the technological boundaries, and the core semantic instructions are used to determine the dimensions of innovation assessment. Based on the constraint semantic vector, a set of benchmark patents that are within the same technical boundary as the target patent is determined; Based on the core semantic vectors of each benchmark patent in the benchmark patent set, a global comparison benchmark vector is determined. Based on the core semantic vector of the target patent and the global comparison benchmark vector, the degree of innovation of the target patent in the innovation examination dimension is evaluated. The patent dataset includes multiple patents from a specific time period, and these multiple patents belong to the same technical field as the target patent.

2. The evaluation method according to claim 1, characterized in that, The step of determining the set of benchmark patents within the same technical boundary as the target patent based on the constraint semantic vector includes: The benchmark patent set is determined based on the first similarity between the constraint semantic vector corresponding to the target patent and the constraint semantic vector corresponding to each patent in the patent dataset.

3. The evaluation method according to claim 2, characterized in that, Determining the benchmark patent set includes: Calculate the first similarity between the constraint semantic vector corresponding to the target patent and the constraint semantic vector corresponding to each patent in the patent dataset; After sorting the patents by first similarity from high to low, the top k patents with the highest first similarity are selected to form the benchmark patent set, where k is an integer greater than or equal to 2.

4. The evaluation method according to claim 2 or 3, characterized in that, The determination of the global comparison benchmark vector based on the core semantic vectors of each benchmark patent in the benchmark patent set includes: The global comparison benchmark vector is determined based on the first similarity and core semantic vector corresponding to each benchmark patent.

5. The evaluation method according to claim 4, characterized in that, Determining the global comparison benchmark vector includes: The first similarity corresponding to each of the benchmark patents is normalized to obtain the normalized weighted coefficients corresponding to each of the benchmark patents. Based on the normalized weighting coefficients corresponding to each of the benchmark patents, the corresponding core semantic vectors are weighted and fused to obtain a fused vector; The fusion vector is normalized using Euclidean norm to obtain the global comparison benchmark vector.

6. The evaluation method according to any one of claims 1-3 and 5, characterized in that, The evaluation of the innovation level of the target patent in a specified dimension, based on the core semantic vector of the target patent and the global comparison benchmark vector, includes: Calculate the second similarity between the core semantic vector corresponding to the target patent and the global comparison benchmark vector; Based on the second similarity and the preset threshold, the degree of innovation of the target patent in the innovation examination dimension is determined.

7. The evaluation method according to any one of claims 1-3 and 5, characterized in that, Also includes: The instruction control embedding model is trained through the following process: The training dataset and semantic instructions of at least two dimensions are input into the initial instruction embedding model. The model performs forward inference and outputs the semantic instruction embedding vector corresponding to each training sample in the training dataset. Each training sample in the training dataset includes at least: the title and abstract of the patent. Based on the semantic instruction embedding vectors corresponding to each training sample, the third similarity between training samples in the training dataset is determined. Based on the third similarity, a triplet training sample set is constructed; wherein, each triplet training sample in the triplet training sample set includes: anchor sample, positive sample, and negative sample; The training is performed in multiple rounds based on a deep learning framework and an adaptive gradient optimizer. During the training process, the triplet sample set is input into the initial instruction embedding model, and the parameters of the initial instruction embedding model are updated by backpropagation using a joint loss function to obtain the instruction control embedding model. The joint loss function is formed by a weighted fusion of the first loss function and the second loss function; The first loss function represents the instruction-specific contrastive loss, which is used to reduce the vector distance between the anchor sample and the positive sample and increase the vector distance between the anchor sample and the negative sample in the vector space corresponding to a single instruction, so that the model learns the ability to extract subdivided semantics in accordance with the question instruction. The second loss function represents the global embedding regularization contrastive loss, which is used to maintain the distance between anchor samples, positive samples, and negative samples in a general vector space without instructions, so that the model avoids overfitting and retains the ability to generalize to general semantics.

8. The evaluation method according to claim 7, characterized in that, The temperature hyperparameter configured during training is used to adjust the gradient scaling scale for similarity calculation between semantic instruction embedding vectors, in order to balance the convergence speed of the loss function with the discrimination accuracy. The loss function includes the first loss function and the second loss function.

9. The evaluation method according to any one of claims 1-3, 5, and 8, characterized in that, The global comparison benchmark vector is used to characterize the overall technological distribution of benchmark patents in terms of innovation examination dimensions within the same technological boundary.

10. A device for evaluating the degree of patent innovation, characterized in that, include: The vector generation module is used to input the target patent and patent dataset, as well as the constraint semantic instructions and core semantic instructions, into the pre-trained instruction control embedding model to obtain the constraint semantic vector and core semantic vector corresponding to each patent. The constraint semantic instructions are used to determine the technological boundaries, and the core semantic instructions are used to determine the dimensions of innovation assessment. The determination module is used to determine a set of benchmark patents that are within the same technical boundary as the target patent based on the constraint semantic vector; and to determine a global comparison benchmark vector based on the core semantic vectors of each benchmark patent in the set of benchmark patents. The evaluation module is used to evaluate the degree of innovation of the target patent in the innovation examination dimension based on the core semantic vector of the target patent and the global comparison benchmark vector. The patent dataset includes multiple patents from a specific time period, and these multiple patents belong to the same technical field as the target patent.

11. An electronic device comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the evaluation method according to any one of claims 1 to 9.

12. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the evaluation method according to any one of claims 1 to 9.

13. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the evaluation method according to any one of claims 1 to 9.

Images