Patent value evaluation method based on semantic mining and feature fusion

Abstract

The invention discloses a patent value evaluation method based on semantic mining and feature fusion, and the method comprises the following steps: firstly obtaining a field patent data set through retrieval and data cleaning, and defining a patent value label; determining patent value influence factors from two aspects of patent value evaluation indexes and text semantic features; relevant data are obtained and preprocessed; constructing and training a patent value evaluation model comprising four modules of multi-modal data input, feature extraction, feature fusion and patent value prediction; and evaluating and optimizing the performance of the model by adopting a multi-classification evaluation index, and evaluating the value of the target patent by using the optimal model. According to the method, key features in the patent value influence factors can be effectively mined, comprehensive and accurate evaluation of the patent value is realized, and data support and decision basis are provided for patent achievement conversion and management.

Description

Technical Field

[0001] The present invention relates to the technical field of intellectual property and patent value assessment, and in particular to a patent value assessment method based on semantic mining and feature fusion. Background Art

[0002] With the accelerated development of global economic integration and technological progress, intellectual property has become an important part of a country's comprehensive strength. As the core content of intellectual property, strengthening the transformation and application of patents is an inherent need to promote high-quality development, and it is also a necessary part of promoting the construction of a strong intellectual property country. Although the number of patent achievements has increased year by year, the difficulty of value discovery has also increased, resulting in a low transfer conversion rate of patents, and a large number of valid patents are still in a "sleeping" state. Comprehensive and accurate assessment of patent value has become an important means to promote the transformation of patent achievements.

[0003] Based on the types of patent features used, current patent value assessment methods can be divided into four categories:

[0004] 1) Economic method. This method uses the method of evaluating intangible assets in economics to evaluate the patent value, including cost method, market method, income method and real option method. Although the system of this type of evaluation method is relatively mature, the results are relatively rough, random and the parameter estimation of the calculation formula is highly subjective.

[0005] 2) Evaluation method based on scientific measurement. This method starts from the perspective of bibliometrics, constructs a multidimensional patent value evaluation index system, and combines comprehensive evaluation methods (such as analytic hierarchy process AHP, entropy weight method, etc.) or machine learning models (such as support vector machine, random forest) to evaluate the value of patents. Although this type of method balances subjective and objective factors to a certain extent, the indicators and indicator weights involved are different, and they rely too much on patent indicators, and have limited ability to identify nonlinear features of data. In addition, this type of method is also greatly affected by the quality of patent samples and ignores the potential value in patent text information.

[0006] 3) Evaluation method based on text analysis. This method uses natural language processing (NLP) technology to explore the potential value in patent texts, quantify and analyze the technical relevance and semantic features in patent texts, and effectively reveal the deep connotation of patent technology innovation. However, the terms, concepts and technical features in different technical fields vary greatly, and the existing text semantic feature extraction methods are difficult to flexibly apply to cross-field patents, thus affecting the comparability between patents. In addition, relying solely on text analysis cannot fully reflect the multi-dimensional value of patents, which may lead to unstable patent value evaluation results.

[0007] 4) Multi-dimensional feature fusion evaluation method. This method usually uses deep learning and feature fusion technology to unify multi-dimensional features such as patent indicator attributes and text content into the same model, overcoming the shortcomings of single feature evaluation and capturing the value of patents more comprehensively. However, the multi-dimensional feature fusion evaluation method is still in the preliminary research stage. The correlation and redundancy between features may affect the effect of the evaluation model, and the model performance depends on feature fusion technology. At the same time, the method currently lacks relevant visualization design, and it is difficult to determine which parts of the features the model focuses on, which leads to poor model interpretability.

[0008] In summary, the existing patent value assessment methods currently have the following problems: 1) limited ability to identify nonlinear features of patent indicator data; 2) the model’s cross-domain applicability is not strong; 3) there is a lack of technical solutions that can effectively integrate multidimensional and multimodal features; 4) the model’s interpretability is poor. Summary of the invention

[0009] The technical problem to be solved by the present invention is to address the deficiencies of the above-mentioned prior art, and the purpose of the present invention is to propose a patent value assessment method based on semantic mining and feature fusion. The method aims to improve the generalization ability of the model by deeply mining the semantic features of patent texts and combining multimodal feature fusion technology, thereby making patent value assessment more comprehensive and accurate.

[0010] To achieve the above object, the present invention provides the following technical solutions:

[0011] The present invention provides a patent value assessment method based on semantic mining and feature fusion, comprising the following steps:

[0012] Step 1: Define patent value labels according to the purpose and needs of the assessment; then obtain patent information in a specific field or specific subject through patent database retrieval, build a field patent data set, and perform preliminary preprocessing and calculate patent value labels;

[0013] Step 2: Determine the factors affecting patent value, involving two aspects: patent value assessment indicators and text semantic features;

[0014] Step 3: Data acquisition and preprocessing: Obtain the corresponding bibliographic information from the field patent data set according to the factors affecting patent value, and remove invalid and duplicate patents; perform different preprocessing operations on patent indicators and text data, and finally obtain the value assessment indicator vector and text input vector;

[0015] Step 4: Build a patent value assessment model. Based on natural language processing technology and multimodal learning technology, build a deep learning model that includes four modules: multimodal data input, feature extraction, feature fusion, and patent value prediction.

[0016] Step 5: Model evaluation and value assessment results. Use accuracy, macro-average precision, macro-average recall, and macro-average F1 value multi-classification evaluation indicators to evaluate and optimize the performance of the model, and use the optimal model to conduct value assessment on the target patent.

[0017] Furthermore, in the step 1, the patent information includes: relevant information for calculating the value label, basic patent information, patent owner information, patent applicant information, patent technical content, patent legal status, patent claims, patent examination process, patent market data, basic information of cited patents, and basic information of cited patents; preliminary preprocessing includes: eliminating patents with invalid legal status, removing duplicate patents according to title, public announcement number, and application number, filling missing values, and converting time data format; value label setting types include value, patent transaction price, and patent level.

[0018] Furthermore, the step 2 is specifically as follows: based on bibliometric indicators, a patent value evaluation index system is constructed from the legal value dimension, technical value dimension and economic value dimension, which includes 3 first-level indicators, 10 second-level indicators and 30 third-level indicators; semantic features are extracted from the title and abstract text to measure creativity, practicality and novelty from a semantic perspective.

[0019] Furthermore, the step 3 specifically includes the following steps:

[0020] S301: Obtaining bibliographic information: Obtaining relevant bibliographic information from the patent database according to the patent value influencing factors determined in step 2, wherein the bibliographic information corresponding to the patent value evaluation index includes the target patent information, the cited patent information and the cited patent information, and the bibliographic information corresponding to the text semantic feature part includes the title and abstract of the target patent;

[0021] S302: Calculate the index value, and obtain the index value according to the acquired bibliographic information and the index calculation method;

[0022] S303: Standardization of numerical indicators, using the range transformation method to standardize data values ​​to eliminate the impact of dimensions;

[0023] S304: Category indicator coding, using one-hot coding to encode category indicators, the indicators involved include invalid review and applicant type;

[0024] S305: Obtaining a patent value assessment index vector, and obtaining a preprocessed patent value assessment index input vector based on the above technical solution;

[0025] S306: extracting text data, including the title and abstract data of the patent;

[0026] S307: Preprocessing text input, removing invalid text such as URL links, HTML tags and special symbols, and using the BERT model in natural language processing technology to perform word segmentation, remove stop words, vectorize text, and fill long texts.

[0027] S308: Obtain a text input vector and prepare for multimodal data input.

[0028] Furthermore, the step 4 specifically includes the following steps:

[0029] S401: a multimodal data input module, the module receiving an indicator input vector and a text input vector, wherein the indicator input vector includes a numerical indicator vector and a category indicator vector;

[0030] S402: a feature extraction module, which extracts features from input data of different modalities to obtain feature representations of different data;

[0031] S403: Feature fusion module, adopts adaptive weight fusion strategy, designs adaptive feature fusion module based on multi-head self-attention mechanism, dynamically fuses information of two modalities, index table and text, and obtains fused feature vector V 3 ;

[0032] S404: Patent value prediction module, which integrates the feature vector V 3 Input into the patent value prediction module and use the multi-layer perceptron as the patent value predictor.

[0033] Furthermore, the S402 specifically includes:

[0034] For the indicator input vector, firstly, the two types of indicator input vectors are concatenated to form a complete indicator input vector, and then the multi-layer perceptron is used to extract the features of the patent indicator input vector, and finally the features of the indicator are obtained;

[0035] For text input vectors, the BERT model is used to extract text semantic features, and it is set whether to fine-tune the model based on the model performance and data set conditions. Then, the BERT features are dimensionally converted and nonlinearly transformed through a fully connected layer, and the text semantics are finally output.

[0036] Furthermore, in S403, the adaptive feature fusion module is specifically designed as follows:

[0037] First, the indicator features and text features are passed through the fully connected layer and the sigmoid activation function respectively to generate two adaptive weights, namely the indicator feature weight ω 1 and text feature weight ω 2 , the specific calculation formula is as follows:

[0038] ω i =σ(W i V i +b i )

[0039] Among them, W i ∈R 256×1 and b i is a learnable parameter, W i represents the weight matrix of the fully connected layer, b i represents the bias term of the fully connected layer, and σ(·) is

[0040] Then, the indicator feature V 1 and text features V 2 Perform weighted fusion to obtain the intermediate fusion feature V f , the specific calculation formula is as follows:

[0041]

[0042] in, Represents the Hadamard product, that is, element-by-element multiplication operation, and + represents broadcast addition;

[0043] Finally, the intermediate fusion feature V f Input into the multi-head self-attention mechanism in the Transformer model for further fusion to obtain the fusion feature V 3 , and at the same time, Dropout regularization is applied in the multi-head self-attention mechanism to prevent overfitting.

[0044] Furthermore, in S403, the multi-head self-attention mechanism is implemented as follows:

[0045] First, input feature V f After three different linear transformation layers, Query, Key and Value vectors are generated respectively, and the calculation formula is:

[0046] Q=V f W Q , K=V f W K , V = V f W V

[0047] Among them, W Q , W K and W V represents a trainable weight matrix;

[0048] Then, the self-attention output of each head is calculated as:

[0049]

[0050] Among them, head i is the attention output of the i-th head, h is the number of heads, d k is the input dimension;

[0051] Finally, the results obtained from each head are concatenated and the output of the final multi-head self-attention layer is generated through a linear transformation layer, which is the required fusion feature vector V 3 .

[0052] Furthermore, in S404, the patent value predictor is composed of two fully connected layers, and the dimension of the first fully connected layer is the fusion feature vector V 3 The second fully connected layer is used to predict the final value label, which is mapped to the final prediction space through a fully connected layer. The dimension is the number of labels of the prediction task. In addition, ReLU is used as the activation function after each fully connected layer, and Dropout regularization is applied to prevent overfitting.

[0053] Furthermore, the step 5 specifically includes the following steps:

[0054] S501 Model training: Divide the domain patent data set into training set, validation set and test set, and use the training set to train the patent value assessment model by setting the optimizer and loss function;

[0055] S502 Model Optimization: During the model training and verification process, the learning rate scheduler monitors the loss value of the verification set, automatically saves the rounds and parameters of the optimal model, and replaces the previous saves until the accuracy requirements are met or the number of training times is reached, then the training ends;

[0056] S503 Model Evaluation: After model training and optimization are completed, the performance of the model on the test set is evaluated through evaluation indicators to evaluate the accuracy and generalization ability of the model; if the evaluation result does not meet the requirements, the model is optimized again;

[0057] S504 Patent value assessment result output: The optimal model obtained through the above steps is applied to the value assessment of the target patent to generate the final patent value assessment result.

[0058] Based on the above technical solution, the embodiments of the present invention can at least produce the following technical effects:

[0059] The patent value assessment method based on semantic mining and feature fusion provided by the present invention is based on bibliometrics, natural language processing and deep learning technology. From the two perspectives of patent value assessment index system and text semantic features, the patent value influencing factors are comprehensively analyzed and determined, and based on this, a patent value assessment model based on semantic mining and feature fusion is constructed. Compared with other patent value assessment models, it has the following advantages: introducing patent text semantic features to measure patent value, deeply excavating the deep connotation of patent text, and avoiding the incompleteness of single feature assessment; by designing the indicator feature extractor module, the model's nonlinear feature recognition ability for patent indicator data is improved; by fine-tuning the BERT model, the model's applicability to data sets in different fields is improved; combined with the multi-head self-attention mechanism of the Transformer model, a technical solution that can effectively fuse multidimensional and multimodal features is provided, which can achieve complementary advantages of different modal features and increase the expression ability of features. Through these designs, the stability and generalization of the model are enhanced, and a patent value assessment model with higher accuracy is finally obtained. Through the method of the present invention, the accuracy and objectivity of patent value assessment can be improved, the transformation of patent achievements can be promoted, and data support and decision-making basis can be provided for patent management. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the structures shown in these drawings without paying creative work.

[0061] Figure 1 It is a process framework diagram of the present invention;

[0062] Figure 2 This is a model diagram for evaluating the patent value of the present invention;

[0063] Figure 3 This is a diagram of the adaptive feature fusion module of the present invention.

[0064] Figure 4 This is a flow chart of the model evaluation and value assessment results of the present invention. DETAILED DESCRIPTION

[0065] The technical solutions in the embodiments of the present invention will be described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention. In addition, the technical solutions between the various embodiments can be combined with each other, but they must be based on the ability of ordinary technicians in this field to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

[0066] like Figure 1 As shown, a patent value assessment method based on semantic mining and feature fusion is provided, and the method specifically includes:

[0067] Step 1: Define patent value labels based on the purpose and needs of the evaluation; then obtain patent information in a specific field or specific subject through patent database retrieval, build a field patent data set, and perform preliminary preprocessing and calculate patent value labels.

[0068] Among them, patent information includes: relevant information for calculating value labels, basic patent information, patent owner information, patent applicant information, patent technical content, patent legal status, patent claims, patent examination process, patent market data, basic information of cited patents, and basic information of cited patents; preliminary preprocessing includes: eliminating patents with invalid legal status, removing duplicate patents based on title, public announcement number, application number and other information, filling missing values, and converting time data formats; value labels can be set to types including but not limited to value, patent transaction price and patent level.

[0069] In this embodiment, the value degree is a quantitative score of the comprehensive value of the patent, the patent transaction price is a patent market transaction price range determined according to market transaction conditions, and the patent grade is different grades based on patent value, such as high value, medium value and low value; these value label types are selected and set according to actual implementation needs to meet the patent value assessment needs in different fields and application scenarios.

[0070] Step 2: Determine the factors affecting patent value, which involves two aspects: patent value assessment indicators and text semantic features.

[0071] Step 2 is as follows: Based on bibliometric indicators, a patent value assessment index system is constructed from the legal value dimension, technical value dimension and economic value dimension, which includes 3 first-level indicators, 10 second-level indicators and 30 third-level indicators; semantic features are extracted from texts such as titles and abstracts to measure creativity, practicality and novelty from a semantic perspective.

[0072] Specifically, the 10 secondary indicators are patent validity, right stability, scope of right protection, technological advancement, scope of technological applicability, technological independence, development trend in the field of technology, competitive situation, market application scope and patent operation; the 30 tertiary indicators are remaining validity period of patent, number of simple families, number of extended families, number of countries with patent layout, invalidation after review, number of claims, number of independent claims, number of first right words, technological originality, number of cited patents, number of citations, citation rate, scientific relevance, number of IPCs, number of IPCs of citing patents, number of IPCs of cited patents, number of subject classifications, proportion of self-cited patents, technology life cycle, number of patents of patent holders, type of applicants, number of applicants, number of inventors, number of citations of the family, number of cited applicants of the family, number of national economic classifications, number of national economic industries, transfer and licensing, financing and insurance, and litigation and arbitration. See Table 1 for details.

[0073] Table 1 Patent value assessment index system

[0074]

[0075]

[0076] Step 3: Data acquisition and preprocessing: obtain the corresponding bibliographic information from the field patent data set according to the factors affecting patent value, and remove invalid and duplicate patents; perform different preprocessing operations on patent indicators and text data, and finally obtain the value assessment indicator vector and text input vector.

[0077] Step 3 is as follows:

[0078] S301: Obtain bibliographic information. According to the patent value influencing factors determined in step 2, relevant bibliographic information is obtained from the patent database, wherein the bibliographic information corresponding to the patent value evaluation index includes the target patent information, the cited patent information and the cited patent information, and the bibliographic information corresponding to the text semantic feature part includes but is not limited to the title and abstract of the target patent;

[0079] S302: Calculate the index value, and obtain the index value according to the acquired bibliographic information and the index calculation method;

[0080] The rules for obtaining bibliographic information and calculating indicators are detailed in Table 2.

[0081] Table 2 Obtaining bibliographic information

[0082]

[0083]

[0084]

[0085] The invalidation after review is classified into four types according to the content of the review decision: patent that has undergone review and is valid, patent that has undergone review and is partially valid, patent that has undergone review and is declared invalid, and invalidation without review. The applicant type is further classified into six types according to the bibliographic information: scientific research unit, institutional cooperation, enterprise, school, individual and others.

[0086] The calculation formula for technical originality is:

[0087]

[0088] Among them, Originality i Indicates the technical originality of target patent i, C ij represents the number of IPC subclasses of the jth cited patent of target patent i, TC i represents the sum of IPC subclasses of all cited patents of target patent i, and N represents the total number of patents cited by the target patent.

[0089] The calculation formula for technology life cycle is:

[0090] TCT i =Median(Year i -Year j ), j = 1, 2, 3, ... N

[0091] Among them, TCT i Indicates the life cycle of the technology represented by target patent i, expressed as the median age of all cited patents, Year i Indicates the authorization year of target patent i, Year j represents the authorization year of the jth cited patent of target patent i, and N represents the total number of patents cited by the target patent.

[0092] S303: Standardization of numerical indicators. The range transformation method is used to standardize the data values ​​to eliminate the dimension effect. The specific calculation formula is:

[0093] If the jth indicator is a positive indicator (the larger the better)

[0094]

[0095] If the jth indicator is a negative indicator (the smaller the better)

[0096]

[0097] Among them, max(x j ) is the maximum value of the j-th indicator data, min(x j ) is the minimum value in the j-th indicator data.

[0098] S304: Category indicator coding, using one-hot coding to encode category indicators, the indicators involved include invalid review and applicant type;

[0099] S305: Obtaining a patent value assessment index vector, and obtaining a preprocessed patent value assessment index input vector based on the above technical solution;

[0100] S306: extracting text data, including the title and abstract data of the patent;

[0101] S307: Preprocessing text input, removing invalid text such as URL links, HTML tags and special symbols, and using the BERT model in natural language processing technology to perform word segmentation, remove stop words, vectorize text, and fill long texts.

[0102] S308: Obtain a text input vector and prepare for multimodal data input.

[0103] Step 4: Build a patent value assessment model. Based on natural language processing technology and multimodal learning technology, build a deep learning model that includes four modules: multimodal data input, feature extraction, feature fusion, and patent value prediction.

[0104] like Figure 2 As shown, step 4 is specifically as follows:

[0105] S401: a multimodal data input module, which receives an indicator input vector and a text input vector obtained based on the above technical solution, wherein the indicator input vector includes a numerical indicator vector and a category indicator vector;

[0106] S402: a feature extraction module, which extracts features from input data of different modalities to obtain feature representations of different data;

[0107] In this embodiment, for the indicator input vector, the two types of indicator input vectors are firstly concatenated to form a complete indicator input vector, and then a multi-layer perceptron (MLP) is used to extract features of the patent indicator input vector, and finally the feature representation V of the indicator is obtained. 1 .

[0108] In this embodiment, the indicator feature extractor consists of three fully connected layers (FC1, FC2 and FC3). The third fully connected layer is used to output the indicator feature vector. The dimensions of each fully connected layer are 64, 128 and 256 respectively. ReLU is used as the activation function after each fully connected layer, and Dropout regularization is applied to prevent overfitting. See Table 3 for details.

[0109] In this embodiment, for the text input vector, the BERT model is first used to extract text semantic features, and it is set whether to perform fine-tuning according to the model performance and data set conditions. Then, a fully connected layer (FC4) is used to perform dimension conversion and nonlinear transformation on the BERT features, and finally the text semantic feature representation V is output. 2 , see Table 3 for details.

[0110] Table 3. Detailed information of feature extractor

[0111]

[0112] S403: Feature fusion module, adopts adaptive weight fusion strategy, designs adaptive feature fusion module based on multi-head self-attention mechanism, dynamically fuses information of two modalities, index table and text, and obtains fused feature vector V 3 ;

[0113] like Figure 3 As shown in the figure, the specific design of the adaptive feature fusion module is as follows:

[0114] First, the indicator feature and text feature are passed through a fully connected layer (FC5, FC6) and a sigmoid activation function to generate two adaptive weights, namely the indicator feature weight ω 1 and text feature weight ω 2 , the specific calculation formula is as follows:

[0115] ω i =σ(W i V i +b i )

[0116] Among them, W i ∈R 256×1 and b i is a learnable parameter, W i represents the weight matrix of the fully connected layer, b i represents the bias term of the fully connected layer, and σ(·) is

[0117] Then, the indicator feature V 1 and text features V 2 Perform weighted fusion to obtain the intermediate fusion feature V f , the specific calculation formula is as follows:

[0118]

[0119] in, represents the Hadamard product, which is an element-by-element multiplication operation, and + represents broadcast addition.

[0120] Finally, the intermediate fusion feature Vf Input into the multi-head self-attention mechanism in the Transformer model for further fusion to obtain the fusion feature V 3 , and at the same time, Dropout regularization is applied in the multi-head self-attention mechanism to prevent overfitting.

[0121] The implementation of the multi-head self-attention mechanism is as follows:

[0122] First, input feature V f After three different linear transformation layers, Query, Key and Value vectors are generated respectively, and the calculation formula is:

[0123] QV f W Q ,KV f W K , VV f W V

[0124] Among them, W Q , W K and W V Represents a trainable weight matrix.

[0125] Then, the self-attention output of each head is calculated as:

[0126]

[0127] Among them, head i is the attention output of the i-th head, h is the number of heads, d k is the input dimension.

[0128] Finally, the results obtained from each head are concatenated and the output of the final multi-head self-attention layer is generated through a linear transformation layer, which is the required fusion feature vector V 3 .

[0129] S404: Patent value prediction module, which integrates the feature vector V 3 Input into the patent value prediction module and use the multi-layer perceptron as the patent value predictor.

[0130] In this embodiment, the patent value predictor consists of two fully connected layers. The dimension of the first fully connected layer is the fusion feature vector V 3 The second fully connected layer is used to predict the final value label, which is mapped to the final prediction space through a fully connected layer. The dimension is the number of labels of the prediction task. In addition, ReLU is used as the activation function after each fully connected layer, and Dropout regularization is applied to prevent overfitting.

[0131] Step 5: Model evaluation and value assessment results. Use multiple classification evaluation indicators such as accuracy, macro-average precision, macro-average recall, and macro-average F1 value to evaluate and optimize the performance of the model, and use the optimal model to conduct value assessment on the target patent.

[0132] like Figure 4 As shown, step 5 is specifically as follows:

[0133] S501 Model training: Divide the field patent data set into training set, validation set and test set, and use the training set to train the patent value assessment model by setting the optimizer and loss function. The loss function is set to the cross entropy loss function, and the optimizer uses the Adam optimizer;

[0134] S502 Model Optimization: During the model training and verification process, the learning rate scheduler monitors the loss value of the verification set, automatically saves the rounds and parameters of the optimal model, and replaces the previous saves until the accuracy requirements are met or the number of training times is reached, then the training ends;

[0135] S503 Model Evaluation: After model training and optimization are completed, the performance of the model on the test set is evaluated through evaluation indicators to evaluate the accuracy and generalization ability of the model; if the evaluation result does not meet the requirements, the model is optimized again;

[0136] S504 Patent value assessment result output: The optimal model obtained through the above steps is applied to the value assessment of the target patent to generate the final patent value assessment result.

[0137] The above shows and describes the basic principles and main features of the present invention and the advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, and these changes and improvements fall within the scope of the present invention to be protected. The scope of the present invention to be protected is defined by the attached claims and their equivalents.

Claims

1. A patent value assessment method based on semantic mining and feature fusion, characterized in that: The following steps are involved: Step 1: Define patent value labels according to the purpose and needs of the assessment; then obtain patent information in a specific field or specific subject through patent database retrieval, build a field patent data set, and perform preliminary preprocessing and calculate patent value labels; Step 2: Determine the factors affecting patent value, involving two aspects: patent value assessment indicators and text semantic features; Step 3: Data acquisition and preprocessing: Obtain the corresponding bibliographic information from the field patent data set according to the factors affecting patent value, and remove invalid and duplicate patents; perform different preprocessing operations on patent indicators and text data, and finally obtain the value assessment indicator vector and text input vector; Step 4: Build a patent value assessment model. Based on natural language processing technology and multimodal learning technology, build a deep learning model that includes four modules: multimodal data input, feature extraction, feature fusion, and patent value prediction. Step 5: Model evaluation and value assessment results. Use accuracy, macro-average precision, macro-average recall, and macro-average F1 value multi-classification evaluation indicators to evaluate and optimize the performance of the model, and use the optimal model to conduct value assessment on the target patent.

2. The patent value assessment method based on semantic mining and feature fusion according to claim 1 is characterized in that: In step 1, the patent information includes: relevant information for calculating the value label, basic patent information, patent owner information, patent applicant information, patent technical content, patent legal status, patent claims, patent examination process, patent market data, basic information of cited patents, and basic information of cited patents; preliminary preprocessing includes: eliminating patents with invalid legal status, removing duplicate patents according to title, public announcement number, and application number, filling missing values, and converting time data format; value label setting types include value, patent transaction price, and patent level.

3. The patent value assessment method based on semantic mining and feature fusion according to claim 1 is characterized in that: The step 2 is specifically as follows: based on bibliometric indicators, a patent value assessment index system is constructed from the legal value dimension, technical value dimension and economic value dimension, which includes 3 first-level indicators, 10 second-level indicators and 30 third-level indicators; semantic features are extracted from the title and abstract text to measure creativity, practicality and novelty from a semantic perspective.

4. The patent value assessment method based on semantic mining and feature fusion according to claim 1 is characterized in that: The step 3 specifically comprises the following steps: S301: Obtaining bibliographic information: Obtaining relevant bibliographic information from the patent database according to the patent value influencing factors determined in step 2, wherein the bibliographic information corresponding to the patent value evaluation index includes the target patent information, the cited patent information and the cited patent information, and the bibliographic information corresponding to the text semantic feature part includes the title and abstract of the target patent; S302: Calculate the index value, and obtain the index value according to the acquired bibliographic information and the index calculation method; S303: Standardization of numerical indicators, using the range transformation method to standardize data values ​​to eliminate the impact of dimensions; S304: Category indicator coding, using one-hot coding to encode category indicators, the indicators involved include invalid review and applicant type; S305: Obtaining a patent value assessment index vector, and obtaining a preprocessed patent value assessment index input vector based on the above technical solution; S306: extracting text data, including the title and abstract data of the patent; S307: Preprocessing text input, removing invalid text such as URL links, HTML tags and special symbols, and using the BERT model in natural language processing technology to perform word segmentation, remove stop words, vectorize text, and fill long texts. S308: Obtain a text input vector and prepare for multimodal data input.

5. The patent value assessment method based on semantic mining and feature fusion according to claim 1 is characterized in that: The step 4 specifically comprises the following steps: S401: a multimodal data input module, the module receiving an indicator input vector and a text input vector, wherein the indicator input vector includes a numerical indicator vector and a category indicator vector; S402: a feature extraction module, which extracts features from input data of different modalities to obtain feature representations of different data; S403: feature fusion module, adopts adaptive weight fusion strategy, designs adaptive feature fusion module based on multi-head self-attention mechanism, dynamically fuses information of two modalities, index table and text, and obtains fused feature vector V3; S404: Patent value prediction module, inputs the fused feature vector V3 into the patent value prediction module, and uses a multi-layer perceptron as a patent value predictor.

6. The patent value assessment method based on semantic mining and feature fusion according to claim 5 is characterized in that: The S402 specifically includes: For the indicator input vector, firstly, the two types of indicator input vectors are concatenated to form a complete indicator input vector, and then the multi-layer perceptron is used to extract the features of the patent indicator input vector, and finally the features of the indicator are obtained; For text input vectors, the BERT model is used to extract text semantic features, and it is set whether to fine-tune the model based on the model performance and data set conditions. Then, the BERT features are dimensionally converted and nonlinearly transformed through a fully connected layer, and the text semantics are finally output.

7. The patent value assessment method based on semantic mining and feature fusion according to claim 5 is characterized in that: In S403, the adaptive feature fusion module is specifically designed as follows: First, the indicator features and text features are passed through the fully connected layer and the sigmoid activation function respectively to generate two adaptive weights, namely the indicator feature weight ω1 and the text feature weight ω2. The specific calculation formula is as follows: oh i =σ(W i V i +b i ) Among them, W i ∈R 256×1 and b i is a learnable parameter, W i represents the weight matrix of the fully connected layer, b i represents the bias term of the fully connected layer, and σ(·) is Then, the indicator feature V1 and the text feature V2 are weightedly fused to obtain the intermediate fusion feature V f , the specific calculation formula is as follows: in, Represents the Hadamard product, that is, element-by-element multiplication operation, and + represents broadcast addition; Finally, the intermediate fusion feature V f The input is further fused into the multi-head self-attention mechanism in the Transformer model to obtain the fused feature V3. At the same time, Dropout regularization is applied in the multi-head self-attention mechanism to prevent overfitting.

8. The patent value assessment method based on semantic mining and feature fusion according to claim 5 is characterized in that: In S403, the multi-head self-attention mechanism is implemented as follows: First, input feature V f After three different linear transformation layers, Query, Key and Value vectors are generated respectively, and the calculation formula is: Q-V f W Q ,K-V f W K ,V-V f W V Among them, W Q , W K and W V represents a trainable weight matrix; Then, the self-attention output of each head is calculated as: head t =Attention(QW i Q ,KW i K ,VW i V ),i=1,2,...h Among them, head i is the attention output of the i-th head, h is the number of heads, d k is the input dimension; Finally, the results obtained from each head are concatenated, and the output of the final multi-head self-attention layer is generated through a linear transformation layer, which is the required fused feature vector V3.

9. The patent value assessment method based on semantic mining and feature fusion according to claim 5 is characterized in that: In S404, the patent value predictor consists of two fully connected layers. The dimension of the first fully connected layer is the dimension of the fused feature vector V3. The second fully connected layer is used to predict the final value label, which is mapped to the final prediction space through a fully connected layer. The dimension is the number of labels of the prediction task. In addition, ReLU is used as the activation function after each fully connected layer, and Dropout regularization is applied to prevent overfitting.

10. The patent value assessment method based on semantic mining and feature fusion according to claim 1 is characterized in that: The step 5 specifically comprises the following steps: S501 Model training: Divide the domain patent data set into training set, validation set and test set, and use the training set to train the patent value assessment model by setting the optimizer and loss function; S502 Model Optimization: During the model training and verification process, the learning rate scheduler monitors the loss value of the verification set, automatically saves the rounds and parameters of the optimal model, and replaces the previous saves until the accuracy requirements are met or the number of training times is reached, then the training ends; S503 Model Evaluation: After model training and optimization are completed, the performance of the model on the test set is evaluated through evaluation indicators to evaluate the accuracy and generalization ability of the model; if the evaluation result does not meet the requirements, the model is optimized again; S504 Patent value assessment result output: The optimal model obtained through the above steps is applied to the value assessment of the target patent to generate the final patent value assessment result.

Images