Patent value evaluation method and system based on deep learning

Abstract

The invention discloses a patent value evaluation method and system based on deep learning, and relates to the technical field of patent evaluation, and the method comprises the following steps: obtaining a patent index data set; extracting semantic features of the patent index data set by using a pre-trained large language model, and generating a text vector; performing index completion on the unlabeled patent data set based on a semi-supervised learning normal form, and constructing a training set and a test set; constructing an evaluation model based on deep learning, performing iterative training on the evaluation model by using the training set, and performing evaluation based on the test set and a preset performance evaluation index to obtain a trained evaluation model; obtaining to-be-evaluated patent index data, and inputting the to-be-evaluated patent index data into the trained evaluation model to obtain a prediction score of the to-be-evaluated patent index data; and based on the prediction score, obtaining an evaluation result of the patent index data to be evaluated.

Description

Technical Field

[0001] This application relates to the technical field of patent evaluation, and mainly relates to a patent value evaluation method and system based on deep learning. Background Art

[0002] With the rapid development of artificial intelligence technology, patent value evaluation methods are also evolving continuously. Patent value evaluation is of great significance for an enterprise's technology layout, market competition strategy, and the formulation of the government's science and technology innovation policy. However, traditional patent value evaluation methods mainly rely on expert review, citation analysis, or simple statistical indicators, and have problems such as strong subjectivity, low efficiency, and difficulty in processing large-scale data; especially in the context of the rapid growth of the global patent volume, the manual evaluation method has been difficult to meet the requirements of efficient and accurate evaluation.

[0003] Currently, existing methods are difficult to meet the requirements of the actual scenario of enterprise patent value evaluation. The main reasons are as follows: Existing methods mostly focus on small-scale models in a single technical dimension, such as decision trees, support vector machines, random forests, etc. However, these methods rely on structured data preprocessed manually and are difficult to directly mine the deep semantic information in patent texts (such as abstracts, claims); at the same time, a large amount of patent data in the actual scenario lacks manually labeled tags, resulting in limited generalization ability of the supervised learning model due to insufficient training data; on the other hand, in the actual scenario, due to the high cost of manual annotation, some patent data lacks tags and cannot be directly used for training and prediction; Chinese Patent with Publication No. CN108416535A discloses a method for patent value evaluation based on deep learning, including: "obtaining the attribute features and citation relationships of a patent, and obtaining the corresponding attribute network representation of the patent through the method of attribute network representation; obtaining the text information of the patent, and obtaining the corresponding text representation of the patent through a multi-layer convolutional neural network; using the corresponding attribute network representation and text representation of the patent to evaluate and predict the value of the patent", but this method only obtains the attribute features, citation relationships, and text information of the patent, and does not make full use of these key information that may affect the patent value, which may lead to an incomplete evaluation result; in addition, this method obtains the corresponding attribute network representation of the patent through the method of attribute network representation, and this method is difficult to capture the complex non-linear relationships between patent attributes, affecting the accuracy of the evaluation; at the same time, this method does not consider the situation where part of the input data labels are missing in the real scenario, and the comprehensive utilization of labeled data and unlabeled data is not efficient enough; in addition, this method uses a relatively simple neural network model to extract text information instead of a large language model with strong expression ability, which will lead to insufficient and inaccurate information mining of the patent text content. Summary of the Invention

[0004] To solve the above problems existing in the prior art, the present application provides a patent value evaluation method and system based on deep learning.

[0005] The technical solution of the present application is as follows: On the one hand, the present invention proposes a patent value evaluation method based on deep learning, and the method includes: Obtain a patent index dataset, including a labeled patent dataset and an unlabeled patent dataset, where the patent indexes include manual indexes and machine indexes; the labeled patent dataset is specifically a patent containing manual indexes and machine indexes, and the unlabeled patent dataset is specifically a patent containing machine indexes and lacking manual indexes; Use a pre-trained large language model to extract the semantic features of the patent index dataset to generate text vectors; complete the indexes of the unlabeled patent dataset based on the semi-supervised learning paradigm, and construct a training set and a test set; Construct an evaluation model based on deep learning, iteratively train the evaluation model using the training set, and evaluate based on the test set and preset performance evaluation indexes to obtain a trained evaluation model; Obtain the patent index data to be evaluated, input the patent index data to be evaluated into the trained evaluation model to obtain the predicted score of the patent index data to be evaluated; based on the predicted score, obtain the evaluation result of the patent index data to be evaluated.

[0006] Preferably, the method further includes data cleaning of the patent index dataset, and the data cleaning includes processing missing values, outliers, and data format standardization.

[0007] Preferably, using a pre-trained large language model to extract the semantic features of the patent index dataset to generate text vectors, which is expressed by the formula: ; ; In the formula, represents the text vector of the th labeled patent data; represents the text vector of the th unlabeled patent data; represents a preset text prompt; represents the th text content of the labeled patent data; represents the th text content of the unlabeled patent data; represents a preset sentence embedding model; represents a pre-trained large language model; represents a concatenation operation; Indicates the index value of the th labeled patent data; Indicates the index value of the th unlabeled patent data.

[0008] Preferably, the metrics of the unlabeled patent dataset are complemented based on the semi-supervised learning paradigm, specifically: Calculate the text similarity between the text vectors of the labeled patent data and the text vectors of the unlabeled patent data, which is expressed by the formula: ; In the formula, Indicates the th text vector of the labeled patent data and the th text vector of the unlabeled patent data; Arrange the generated text similarities in descending order, and select the artificial metrics of the first labeled patent data to form an artificial metric set denoted as , where represents the preset number selected; Based on the artificial metric set, complement the artificial metrics of the unlabeled patent data, and generate the pseudo-labels of the unlabeled patent data based on the artificial metrics, which is expressed by the formula: ; ; ; In the formula, Indicates the th complemented artificial metric of the unlabeled patent data; Indicates the th artificial metric of the labeled patent data; Indicates the th pseudo-label of the unlabeled patent data based on the artificial metrics; Indicates the th value score of the labeled patent data; Indicates the th confidence level of the pseudo-label of the unlabeled patent data based on the artificial metrics; Calculate the machine metric similarity between the machine metrics of the labeled patent data and the machine metrics of the unlabeled patent data, which is expressed by the formula: ; In the formula, Indicates the th machine metric of the labeled patent data and the The similarity of machine metrics between unlabeled patent data Arrange the generated machine metric similarities in descending order and select the first machine metrics of labeled patent data to form a machine metric set denoted as ; Based on the machine metric set, complete the machine metrics of unlabeled patent data and generate pseudo-labels of unlabeled patent data based on machine metrics, which is expressed by the formula: ; ; In the formula, represents the pseudo-label of the th unlabeled patent data based on machine metrics; represents the confidence of the pseudo-label of the th unlabeled patent data based on machine metrics; Perform weighted calculation on the pseudo-labels based on manual metrics and machine metrics and their corresponding confidences to obtain the final pseudo-labels and final confidences of unlabeled patent data, which is expressed by the formula: ; ; In the formula, represents the final pseudo-label of the th unlabeled patent data; represents the final confidence of the th unlabeled patent data; represents the preset index weight.

[0009] Preferably, the specific steps for constructing the training set and the test set are as follows: According to the preset confidence threshold, screen the high-confidence set , where represents the preset confidence threshold; The training set is specifically the combination of the high-confidence set and the labeled patent data set, denoted as , where represents the labeled patent data set; The test set is specifically the unlabeled patent data set after index completion, denoted as .

[0010] On the other hand, the present invention also proposes a patent value evaluation system based on deep learning. The system includes a data acquisition module, an index completion module, a patent evaluation module, and a result output module, where: The data acquisition module is used to acquire a patent index dataset, including a labeled patent dataset and an unlabeled patent dataset, where the patent indexes include manual indexes and machine indexes; the labeled patent dataset specifically refers to patents containing manual indexes and machine indexes, and the unlabeled patent dataset specifically refers to patents containing machine indexes but lacking manual indexes; the patent index dataset is transmitted to the index completion module; The index completion module is used to extract the semantic features of the patent index dataset by using a pre-trained large language model to generate text vectors; perform index completion on the unlabeled patent dataset based on a semi-supervised learning paradigm, and construct a training set and a test set; The patent evaluation module is used to construct an evaluation model based on deep learning, iteratively train the evaluation model by using the training set, and perform evaluation based on the test set and preset performance evaluation indexes to obtain a trained evaluation model; Acquire the patent index data to be evaluated, input the patent index data to be evaluated into the trained evaluation model to obtain the predicted score of the patent index data to be evaluated; based on the predicted score, obtain the evaluation result of the patent index data to be evaluated; The result output module is used to display the evaluation result of the patent index data to be evaluated.

[0011] Preferably, the semantic features of the patent index dataset are extracted by using a pre-trained large language model to generate text vectors, which is expressed by the formula: ; ; In the formula, represents the text vector of the th labeled patent data; represents the text vector of the th unlabeled patent data; represents a preset text prompt; represents the th text content of the labeled patent data; represents the th text content of the unlabeled patent data; represents a preset sentence embedding model; represents a pre-trained large language model; represents a concatenation operation; represents the index value of the th labeled patent data; represents the index value of the th unlabeled patent data.

[0012] Preferably, the metrics of the unlabeled patent dataset are complemented based on the semi-supervised learning paradigm, and the training set and the test set are constructed as follows: Calculate the text similarity between the text vectors of the labeled patent data and the text vectors of the unlabeled patent data, which is expressed by the formula: ; In the formula, represents the text similarity between the text vector of the th labeled patent data and the text vector of the th unlabeled patent data; Arrange the generated text similarities in descending order, and select the artificial metrics of the first labeled patent data to form an artificial metric set denoted as , where represents the preset number selected; Based on the artificial metric set, complement the artificial metrics of the unlabeled patent data to generate pseudo-labels of the unlabeled patent data based on the artificial metrics, which is expressed by the formula: ; ; ; In the formula, represents the complemented artificial metric of the th unlabeled patent data; represents the artificial metric of the th labeled patent data; represents the pseudo-label of the th unlabeled patent data based on the artificial metrics; represents the value score of the th labeled patent data; represents the confidence level of the pseudo-label of the th unlabeled patent data based on the artificial metrics; Calculate the machine metric similarity between the machine metrics of the labeled patent data and the machine metrics of the unlabeled patent data, which is expressed by the formula: ; In the formula, represents the machine metric similarity between the machine metric of the th labeled patent data and the machine metric of the th unlabeled patent data; Arrange the generated machine metric similarities in descending order, and select the machine metrics of the first labeled patent data to form a machine metric set denoted as ; Complete the machine metrics of the unlabeled patent data based on the set of machine metrics, and generate pseudo-labels of the unlabeled patent data based on the machine metrics, which is expressed by the formula: ; ; In the formula, represents the pseudo-label of the th unlabeled patent data based on the machine metrics; represents the confidence of the pseudo-label of the th unlabeled patent data based on the machine metrics; Perform weighted calculation on the pseudo-labels based on artificial metrics and machine metrics and their corresponding confidences to obtain the final pseudo-label and final confidence of the unlabeled patent data, which is expressed by the formula: ; ; In the formula, represents the final pseudo-label of the th unlabeled patent data; represents the final confidence of the th unlabeled patent data; represents the preset metric weight; According to the preset confidence threshold, screen the high-confidence set , where represents the preset confidence threshold; The training set is specifically represented as the combination of the high-confidence set and the labeled patent data set as , where represents the labeled patent data set; The test set is specifically the unlabeled patent data set after metric completion, which is represented as .

[0013] On the other hand, the present invention also proposes an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, it implements a patent value evaluation method based on deep learning as described in the embodiments of the present invention.

[0014] On the other hand, the present invention also proposes a computer-readable storage medium, on which a computer program is stored. When the program is executed by a processor, it implements a patent value evaluation method based on deep learning as described in the embodiments of the present invention.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1) The present invention provides a method and system for patent value evaluation based on deep learning, which uses a pre-trained large language model to extract semantic features of a patent index dataset and generate text vectors, improving the depth of mining patent text information and being able to more accurately capture deep semantic information in the text; 2) The present invention provides a method and system for patent value evaluation based on deep learning. Through a semi-supervised learning paradigm, the overall utilization efficiency of patent data is improved, and index completion is performed on the unlabeled patent dataset, increasing the scale and diversity of training data and enhancing the generalization ability of the model; 3) The present invention provides a method and system for patent value evaluation based on deep learning. An evaluation model based on deep learning is constructed, enhancing the model's ability to handle complex non-linear relationships and improving the accuracy and reliability of the evaluation results; Through iterative training and an evaluation mechanism based on a test set, the model parameters are continuously optimized, enhancing the stability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 is a flowchart of the method of an embodiment of the present invention; Figure 2 is a schematic diagram showing the influence of index weights on the expert mixture model of an embodiment of the present invention; Figure 3 is a schematic diagram showing the influence of confidence index weights on the expert mixture model of an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0017] The following describes the specific embodiments of the present invention to facilitate those skilled in the art of the present technology to understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those of ordinary skill in the art of the present technology, as long as various changes are within the spirit and scope of the present invention defined and determined by the appended claims, these changes are obvious, and all inventions created using the concept of the present invention are within the scope of protection.

[0018] The present invention provides the following technical solution: A method and system for patent value evaluation based on deep learning.

[0019] Embodiment 1: Specifically refer to Figure 1 , this embodiment provides a method for patent value evaluation based on deep learning. The specific steps include: S1. Obtain a patent index dataset, including a labeled patent dataset and an unlabeled patent dataset, where the patent indexes include manual indexes and machine indexes; The labeled patent dataset is specifically a patent containing manual indexes and machine indexes, and the unlabeled patent dataset is specifically a patent containing machine indexes and lacking manual indexes; In this embodiment, the patent indicator data set includes 10,000 pieces of patent data, of which the annotated patent data set is 5,000 pieces of patent data and the unannotated patent data set is 5,000 pieces of patent data; each piece of patent data is composed of 20 indicators, of which 9 are manual indicators and 11 are machine indicators; The artificial indicators include innovation, social benefits, application prospects, technology maturity, industry influence, policy compliance, technology compatibility, international competitiveness and industry chain driving force; The machine indicators include patent life, number of right holders, number of patent families, number of patent lawsuits, patent authorization time, number of patent citations, number of patent conversions, patent ranking, number of patent-related projects, patent authorization time and market share of patent technology applications; The method further includes performing data cleaning on the patent indicator data set, wherein the data cleaning includes processing missing values, outliers and unifying data formats; S2. Use the pre-trained large language model to extract the semantic features of the patent indicator dataset and generate a text vector, which is expressed as: ; ; In the formula, Indicates text vectors of annotated patent data; Indicates Text vectors of unlabeled patent data; Indicates the preset text prompt word; Indicates The text content of the annotated patent data; Indicates Text content of unannotated patent data; Represents the preset sentence embedding model; Represents a pre-trained large language model; Indicates a connection operation; Indicates Index value of annotated patent data; Indicates Index value of unannotated patent data; In this embodiment, the pre-trained large language model For the ChatGLM3-6B model, the preset sentence embedding model It is the text2vec-base-chinese model; S3, complete the indicators of unlabeled patent dataset based on semi-supervised learning paradigm; Calculate the text similarity between the text vectors of the labeled patent data and the unlabeled patent data, which is expressed by the formula: ; In the formula, represents the text similarity between the text vector of the -th labeled patent data and the text vector of the -th unlabeled patent data; Arrange the generated text similarities in descending order, and select the artificial indicators of the first labeled patent data to form an artificial indicator set denoted as , where represents the preset number selected; Based on the artificial indicator set, complete the artificial indicators of the unlabeled patent data, and generate the pseudo-labels of the unlabeled patent data based on the artificial indicators, which is expressed by the formula: ; ; ; In the formula, represents the completed artificial indicator of the -th unlabeled patent data; represents the artificial indicator of the -th labeled patent data; represents the pseudo-label of the -th unlabeled patent data based on the artificial indicator; represents the value score of the -th labeled patent data; represents the confidence of the pseudo-label of the -th unlabeled patent data based on the artificial indicator; Calculate the machine indicator similarity between the machine indicators of the labeled patent data and the machine indicators of the unlabeled patent data, which is expressed by the formula ; In the formula, represents the machine indicator similarity between the machine indicator of the -th labeled patent data and the machine indicator of the -th unlabeled patent data; Arrange the generated machine indicator similarities in descending order, and select the machine indicators of the first labeled patent data to form a machine indicator set denoted as ; Complete the machine metrics of the unlabeled patent data based on the set of machine metrics, and generate pseudo-labels of the unlabeled patent data based on the machine metrics, which is expressed by the formula: ; ; In the formula, represents the pseudo-label of the th unlabeled patent data based on the machine metrics; represents the confidence of the pseudo-label of the th unlabeled patent data based on the machine metrics; Perform weighted calculation on the pseudo-labels based on artificial metrics and machine metrics and their corresponding confidences to obtain the final pseudo-label and final confidence of the unlabeled patent data, which is expressed by the formula: ; ; In the formula, represents the final pseudo-label of the th unlabeled patent data; represents the final confidence of the th unlabeled patent data; represents the preset index weight; In this embodiment, the preset index weight is set in [0, 1], and the step size for increase is 0.1; the preset number selected is 4; S4. Construct a training set and a test set, and screen the high-confidence set according to the preset confidence threshold , where represents the preset confidence threshold; The training set is specifically represented as the combination of the high-confidence set and the labeled patent data set , where represents the labeled patent data set; The test set is specifically represented as the unlabeled patent data set after index completion ; In this embodiment, the preset confidence threshold is set in [0.9, 0.95], and the step size for increase is 0.01; S5. Construct an evaluation model based on deep learning, perform iterative training on the evaluation model using the training set, and evaluate based on the test set and the preset performance evaluation metrics to obtain the trained evaluation model; In this embodiment, the evaluation model based on deep learning constructed is an expert mixture model; the expert mixture model includes An expert network and a gating network, where each expert network is specifically composed of two fully connected layers; S51. Iteratively train the expert mixture model using the training set; input the training set into the expert mixture model to obtain the predicted score of the patent value, which is expressed by the formula: , ; ; ; ; In the formula, represents the predicted score of the patent value; represents the weight of the th expert network; represents the th expert network; represents the number of expert networks; represents the th index value of the expert network; represents the activation function; represents the data of the input training set; represents the th weight of the first fully connected layer in the th expert network; represents the th weight of the second fully connected layer in the th expert network; represents the exponential function; S52. Optimize the expert mixture model using the mean squared error loss function, which is expressed by the formula: ; In the formula, represents the mean squared error loss function; represents the pseudo-label of the th training set data; represents the number of data in the training set; represents the th index value of the training set data; S6. And evaluate based on the test set and the preset performance evaluation metrics to obtain the trained expert mixture model; The preset performance evaluation metrics are specifically the mean absolute error, mean absolute percentage error, and root mean square error, which are expressed by the formula: ; ; ; In the formula, represents the mean absolute error; represents the mean absolute percentage error; represents the root mean square error; represents the th predicted score of the patent value for the training set data; represents the th true score of the patent value for the training set data; Please refer to Table 1. In this embodiment, the multi-layer perceptron is used as the baseline method, and the error comparison results of the baseline method, the method combining the baseline method and the semi-supervised learning paradigm, the method combining the baseline method and the expert mixture model, and the method combining the baseline method, the semi-supervised learning paradigm and the expert mixture model under the preset performance evaluation metrics are compared; Table 1 Error Comparison Results

[0020] Furthermore, in this embodiment, the influence of different index weights and confidence thresholds on the performance evaluation metrics of the expert mixture model is also analyzed; please refer to Figure 2 , the index weight When < 0.4, the performance evaluation metrics of the evaluation model are stable; please refer to Figure 3 , the confidence threshold = 0.93, the performance evaluation metrics of the evaluation model are stable; S7. Obtain the patent index data to be evaluated, input the patent index data to be evaluated into the trained evaluation model to obtain the predicted score of the patent index data to be evaluated; based on the predicted score, obtain the evaluation result of the patent index data to be evaluated.

[0021] Embodiment 2: This embodiment provides a patent value evaluation system based on deep learning. The system includes a data acquisition module, an index completion module, a patent evaluation module, and a result output module, where: The data acquisition module is used to acquire a patent index data set, including a labeled patent data set and an unlabeled patent data set, where the patent index includes artificial indexes and machine indexes; the labeled patent data set is specifically a patent containing artificial indexes and machine indexes, and the unlabeled patent data set is specifically a patent containing machine indexes and lacking artificial indexes; the patent index data set is transmitted to the index completion module; The index completion module is used to extract the semantic features of the patent index dataset using a pre-trained large language model to generate text vectors; perform index completion on the unlabeled patent dataset based on a semi-supervised learning paradigm, and construct a training set and a test set. Extract the semantic features of the patent index dataset using a pre-trained large language model to generate text vectors, which can be expressed by the formula: ; ; In the formula, represents the text vector of the th labeled patent data; represents the text vector of the th unlabeled patent data; represents a preset text prompt; represents the th text content of the labeled patent data; represents the th text content of the unlabeled patent data; represents a preset sentence embedding model; represents a pre-trained large language model; represents a concatenation operation; represents the index value of the th labeled patent data; represents the index value of the th unlabeled patent data; Perform index completion on the unlabeled patent dataset based on a semi-supervised learning paradigm, and construct a training set and a test set, specifically: Calculate the text similarity between the text vectors of the labeled patent data and the text vectors of the unlabeled patent data, which can be expressed by the formula: ; In the formula, represents the text similarity between the text vector of the th labeled patent data and the text vector of the th unlabeled patent data; Arrange the generated text similarities in descending order, select the artificial indices of the first labeled patent data to form an artificial index set denoted as , where represents the preset number selected; Based on the artificial index set, complete the artificial indices of the unlabeled patent data to generate pseudo-labels of the unlabeled patent data based on the artificial indices, which can be expressed by the formula: ; ; ; In the formula, represents the th artificial index for completing unlabeled patent data; represents the th artificial index for labeled patent data; represents the th pseudo-label of unlabeled patent data based on artificial indexes; represents the th value score of labeled patent data; represents the th confidence level of the pseudo-label of unlabeled patent data based on artificial indexes; Calculate the machine index similarity between the machine indexes of labeled patent data and unlabeled patent data, which is expressed by the formula: ; In the formula, represents the th machine index similarity between the machine index of the th labeled patent data and the machine index of the th unlabeled patent data; Arrange the generated machine index similarities in descending order, and select the machine indexes of the first ; Based on the machine index set, complete the machine indexes of unlabeled patent data, and generate the pseudo-labels of unlabeled patent data based on machine indexes, which is expressed by the formula: ; ; In the formula, represents the th pseudo-label of unlabeled patent data based on machine indexes; represents the th confidence level of the pseudo-label of unlabeled patent data based on machine indexes; Perform weighted calculation on the pseudo-labels based on artificial indexes and machine indexes and their corresponding confidence levels to obtain the final pseudo-labels and final confidence levels of unlabeled patent data, which is expressed by the formula: ; ; In the formula, represents the th final pseudo-label of unlabeled patent data; represents the The final confidence level of unlabeled patent data; Indicates the preset index weight; According to the preset confidence threshold, filter the high-confidence set , where Indicates the preset confidence threshold; The training set is specifically represented as the combination of the high-confidence set and the labeled patent data set as , where Indicates the labeled patent data set; The test set is specifically represented as the unlabeled patent data set after index completion as ; The patent evaluation module is used to construct an evaluation model based on deep learning, iteratively train the evaluation model using the training set, and evaluate based on the test set and the preset performance evaluation index to obtain a trained evaluation model; Obtain the patent index data to be evaluated, input the patent index data to be evaluated into the trained evaluation model to obtain the predicted score of the patent index data to be evaluated; based on the prediction judgment, obtain the evaluation result of the patent index data to be evaluated; The result output module is used to display the evaluation result of the patent index data to be evaluated.

[0022] Embodiment 3: This embodiment provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, it implements a patent value evaluation method based on deep learning as described in any embodiment of the present invention.

[0023] Embodiment 4: This embodiment provides a computer-readable storage medium, on which a computer program is stored. When the program is executed by a processor, it implements a patent value evaluation method based on deep learning as described in any embodiment of the present invention.

[0024] It should be noted that the system, electronic device, and computer-readable storage medium described in the present invention are all based on the same principle as the method described in Embodiment 1, and will not be elaborated here.

[0025] The above are only the embodiments of the present invention, and do not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the specification and drawings of the present invention, or directly or indirectly applied in other related technical fields, shall be included in the patent protection scope of the present invention by the same token.

Claims

1. A patent value assessment method based on deep learning, characterized in that: The method comprises: Obtaining a patent indicator dataset, including an annotated patent dataset and an unannotated patent dataset, wherein the patent indicators include manual indicators and machine indicators; the annotated patent dataset specifically refers to patents that include manual indicators and machine indicators, and the unannotated patent dataset specifically refers to patents that include machine indicators but lack manual indicators; The semantic features of the patent indicator dataset are extracted using a pre-trained large language model to generate text vectors; the indicators of the unlabeled patent dataset are completed based on a semi-supervised learning paradigm, and a training set and a test set are constructed; Constructing an evaluation model based on deep learning, iteratively training the evaluation model using the training set, and evaluating the evaluation model based on the test set and preset performance evaluation indicators to obtain a trained evaluation model; Obtain patent indicator data to be evaluated, input the patent indicator data to be evaluated into a trained evaluation model, and obtain a predicted score for the patent indicator data to be evaluated; and obtain an evaluation result of the patent indicator data to be evaluated based on the predicted score.

2. A patent value assessment method based on deep learning according to claim 1, characterized in that: The method also includes performing data cleaning on the patent indicator data set, wherein the data cleaning includes processing missing values, outliers and unifying data formats.

3. A patent value assessment method based on deep learning according to claim 1, characterized in that: The semantic features of the patent indicator dataset are extracted using a pre-trained large language model to generate a text vector, which is expressed as follows: ； ； In the formula, Indicates text vectors of annotated patent data; Indicates Text vectors of unlabeled patent data; Indicates the preset text prompt word; Indicates The text content of the annotated patent data; Indicates Text content of unannotated patent data; Represents the preset sentence embedding model; Represents a pre-trained large language model; Indicates a connection operation; Indicates Index value of annotated patent data; Indicates The index value of unlabeled patent data.

4. A patent value assessment method based on deep learning according to claim 3, characterized in that: Based on the semi-supervised learning paradigm, the indicators of the unlabeled patent dataset are completed, specifically: The text similarity between the text vector of the annotated patent data and the text vector of the unannotated patent data is calculated and expressed as: ； In the formula, Indicates The text vector of the annotated patent data is The text similarity between the text vectors of unlabeled patent data; Arrange the generated text similarities in descending order and select the first The artificial indicators of the annotated patent data constitute the artificial indicator set represented as ,in Indicates the number of selected presets; Based on the artificial indicator set, the artificial indicators of the unlabeled patent data are completed to generate pseudo labels of the unlabeled patent data based on artificial indicators, which can be expressed as follows: ； ； ； In the formula, Indicates Artificial indicators to complete unlabeled patent data; Indicates Artificial indicators of annotated patent data; Indicates Pseudo-labeling of unlabeled patent data based on artificial indicators; Indicates Value scores of annotated patent data; Indicates The confidence of pseudo-labels of unlabeled patent data based on artificial indicators; The machine indicator similarity between the machine indicators of the annotated patent data and the machine indicators of the unannotated patent data is calculated and expressed as: ； In the formula, Indicates The machine indicators of the annotated patent data are The machine indicator similarity between the machine indicators of unlabeled patent data; Arrange the generated machine index similarities in descending order and select the first The machine indicators of the annotated patent data constitute the machine indicator set expressed as ; Based on the machine indicator set, the machine indicators of the unlabeled patent data are completed to generate pseudo labels of the unlabeled patent data based on machine indicators, which can be expressed as follows: ； ； In the formula, Indicates Pseudo-labeling of unlabeled patent data based on machine indicators; Indicates The confidence of the pseudo-labels of unlabeled patent data based on machine indicators; The pseudo labels based on manual indicators and machine indicators and their corresponding confidences are weighted to obtain the final pseudo labels and final confidences of the unlabeled patent data, which are expressed as follows: ； ； In the formula, Indicates Final pseudo labels for unlabeled patent data; Indicates The final confidence of unlabeled patent data; Indicates the preset indicator weight.

5. A patent value assessment method based on deep learning according to claim 4, characterized in that: The specific steps to construct the training set and test set are: Filter high confidence sets according to the preset confidence threshold ,in Indicates the preset confidence threshold; The training set is specifically a combination of a high confidence set and annotated patent data set, expressed as ,in represents a collection of annotated patent data; The test set is specifically an unlabeled patent data set after index completion, which is expressed as .

6. A patent value assessment system based on deep learning, characterized in that: The system includes a data acquisition module, an indicator completion module, a patent evaluation module and a result output module, wherein: The data acquisition module is used to acquire a patent indicator data set, including an annotated patent data set and an unannotated patent data set, wherein the patent indicators include manual indicators and machine indicators; the annotated patent data set is specifically a patent containing manual indicators and machine indicators, and the unannotated patent data set is specifically a patent containing machine indicators and lacking manual indicators; the patent indicator data set is transmitted to the indicator completion module; The indicator completion module is used to extract the semantic features of the patent indicator dataset using a pre-trained large language model to generate a text vector; complete the indicators of the unlabeled patent dataset based on a semi-supervised learning paradigm, and construct a training set and a test set; The patent evaluation module is used to construct an evaluation model based on deep learning, iteratively train the evaluation model using the training set, and evaluate it based on the test set and preset performance evaluation indicators to obtain a trained evaluation model; Obtaining patent indicator data to be evaluated, inputting the patent indicator data to be evaluated into a trained evaluation model to obtain a predicted score for the patent indicator data to be evaluated; and obtaining an evaluation result of the patent indicator data to be evaluated based on the predicted score; The result output module is used to display the evaluation results of the patent indicator data to be evaluated.

7. A patent value assessment system based on deep learning according to claim 6, characterized in that: The semantic features of the patent indicator dataset are extracted using a pre-trained large language model to generate a text vector, which is expressed as follows: ； ； In the formula, Indicates text vectors of annotated patent data; Indicates Text vectors of unlabeled patent data; Indicates the preset text prompt word; Indicates The text content of the annotated patent data; Indicates Text content of unannotated patent data; Represents the preset sentence embedding model; Represents a pre-trained large language model; Indicates a connection operation; Indicates Index value of annotated patent data; Indicates The index value of unlabeled patent data.

8. A patent value assessment system based on deep learning according to claim 7, characterized in that: Based on the semi-supervised learning paradigm, the indicators of the unlabeled patent dataset are completed, and the training set and test set are constructed. Specifically: The text similarity between the text vector of the annotated patent data and the text vector of the unannotated patent data is calculated and expressed as: ； In the formula, Indicates The text vector of the annotated patent data is The text similarity between the text vectors of unlabeled patent data; Arrange the generated text similarities in descending order and select the first The artificial indicators of the annotated patent data constitute the artificial indicator set represented as ,in Indicates the number of selected presets; Based on the artificial indicator set, the artificial indicators of the unlabeled patent data are completed to generate pseudo labels of the unlabeled patent data based on artificial indicators, which can be expressed as follows: ； ； ； In the formula, Indicates Artificial indicators to complete unlabeled patent data; Indicates Artificial indicators of annotated patent data; Indicates Pseudo-labeling of unlabeled patent data based on artificial indicators; Indicates Value scores of annotated patent data; Indicates The confidence of pseudo-labels of unlabeled patent data based on artificial indicators; The machine indicator similarity between the machine indicators of the annotated patent data and the machine indicators of the unannotated patent data is calculated and expressed as: ； In the formula, Indicates The machine indicators of the annotated patent data are The machine indicator similarity between the machine indicators of unlabeled patent data; Arrange the generated machine index similarities in descending order and select the first The machine indicators of the annotated patent data constitute the machine indicator set expressed as ; Based on the machine indicator set, the machine indicators of the unlabeled patent data are completed to generate pseudo labels of the unlabeled patent data based on machine indicators, which can be expressed as follows: ； ； In the formula, Indicates Pseudo-labeling of unlabeled patent data based on machine indicators; Indicates The confidence of the pseudo-labels of unlabeled patent data based on machine indicators; The pseudo labels based on manual indicators and machine indicators and their corresponding confidences are weighted to obtain the final pseudo labels and final confidences of the unlabeled patent data, which are expressed as follows: ； ； In the formula, Indicates Final pseudo labels for unlabeled patent data; Indicates The final confidence of unlabeled patent data; Indicates the preset indicator weight; Filter high confidence sets according to the preset confidence threshold ,in Indicates the preset confidence threshold; The training set is specifically a combination of a high confidence set and annotated patent data set, expressed as ,in represents a collection of annotated patent data; The test set is specifically an unlabeled patent data set after index completion, which is expressed as .

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, it implements a patent value assessment method based on deep learning as described in any one of claims 1 to 5.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, a patent value assessment method based on deep learning as described in any one of claims 1 to 5 is implemented.

Images