Pre-application patent quality assessment method and system based on cluster center representation

Abstract

The present invention discloses a patent quality assessment method and system before application based on cluster center representation, including: extracting keywords based on patent text input by users for retrieval, generating sub-datasets with similar features in patent big data, and generating center representation of the sub-datasets through clustering models; intercepting patent information to be predicted in the patent text input by users to generate text representation; calculating the similarity between the text representation and the center representation of the patent information to be predicted, and generating constraint information based on the similarity combined with patent quality indicators; using constraint information to train patent quality assessment models, and obtaining multi-dimensional quality evaluation results for patents input by users. While solving the problem of massive data comparison, the present invention quickly conducts multi-dimensional quality analysis on patents that users plan to apply for, which is conducive to improving the success rate of user applications and cultivating high-value patents, and reducing the cost of patent applications for enterprises.

Description

Technical Field

[0001] The present invention relates to the technical field of patent quality assessment, and more specifically, to a method and system for pre-application patent quality assessment based on cluster center representation. Background Art

[0002] Patents are an important part of intellectual property and the main result of scientific and technological innovation. The number of patents reflects the overall scale of patents, and the quality of patents reflects the quality of patents. At present, the patent level of a region is usually measured by analyzing the number of patents, while the analysis of patent quality is ignored, resulting in a one-sided reflection of the true situation of patents. In recent years, the number of patents has exploded, bringing many challenges to the review and transformation of patents. Therefore, patent quality has received great attention, and the selection of scientific and reasonable patent quality evaluation methods has also become a hot topic in academic research. Especially in the analysis of massive data, constructing segmented small data sets to assist in quality assessment is an important direction for large-scale application of evaluation models.

[0003] At present, the number of patent application documents is increasing rapidly, but the number of patent practitioners is insufficient and their professional knowledge is uneven, which leads to an increase in the workload of patent practitioners and indirectly leads to a decline in the quality of patent application documents. Therefore, patent application documents will affect the quality of patent applications. Improving the quality of patent application documents can fully demonstrate the protection scope of the current enterprise's R&D plan and better provide intellectual property services for enterprises. On the other hand, it can also help improve the quality of patent applications. Therefore, multi-dimensional quality assessment of patent application texts before application is one of the problems that need to be solved urgently. Summary of the invention

[0004] In order to solve the above technical problems, the present invention proposes a pre-application patent quality assessment method and system based on cluster center representation.

[0005] The first aspect of the present invention provides a method for evaluating patent quality before application based on cluster center representation, comprising:

[0006] Extract keywords from patent texts input by users for retrieval, generate sub-datasets whose feature similarity meets preset standards in patent big data, and generate central representations of the sub-datasets through clustering models;

[0007] Extracting the patent information to be predicted from the patent text input by the user, and generating a text representation of the patent information to be predicted;

[0008] Calculate the similarity between the patent information to be predicted and the central representation, and generate constraint information based on the similarity combined with the patent quality index;

[0009] The constraint information is used to train a patent quality assessment model, and a multi-dimensional quality evaluation result is obtained for the patent input by the user through the patent quality assessment model.

[0010] In this solution, keywords are extracted from the patent text input by the user for retrieval, and sub-datasets with similar features that meet the preset standards are generated in the patent big data, specifically:

[0011] Obtain the patent text input by the user for word segmentation preprocessing, generate a serialized representation of the patent text, determine the part-of-speech tags of the word vectors in the serialized representation of the patent text, and use the part-of-speech tags for sequence annotation;

[0012] Using Roberta to trim, segment and embed the serialized representation of the patent text, obtain an embedding vector of the patent text, filter preset phrases through the part-of-speech tags, filter corresponding embedding vectors based on position feature matching of the preset phrases, and splice the matched and filtered embedding vectors to obtain a spliced ​​embedding vector;

[0013] A self-attention mechanism is introduced into the embedding vector of the patent text. The features of the embedding vector are strengthened by weighting the self-attention weight. Cross-attention is introduced between the concatenated embedding vector and the embedded vector to obtain the neighborhood embedding vector of the concatenated embedding vector and strengthen the contextual semantics.

[0014] Obtain the embedded vector sequence and the neighborhood embedded vector sequence after attention encoding, calculate the similarity of the embedded vector and the neighborhood embedded vector in the sequence, obtain the concatenated embedded vector whose similarity meets the preset similarity threshold for decoding, and use it as the keyword extraction result;

[0015] A search index is established based on the keywords, and the keyword feature similarity is calculated in the massive patent big data using the search index to obtain patent data that meets the preset similarity standard to construct a sub-dataset containing the keywords.

[0016] In this solution, the central representation of the sub-dataset is generated by a clustering model, specifically:

[0017] Use the sparrow search algorithm to optimize the initial cluster center of the sub-dataset, initialize the parameters of the sparrow search algorithm, calculate the fitness value in the sparrow population, and obtain the optimal fitness value and the worst fitness value and the corresponding position;

[0018] Select the discoverer, joiner and scout and update the position. In the process of updating the sparrow's position, introduce the adaptive t-distribution variation, iteratively calculate the fitness and update the sparrow's position. After the maximum number of iterations is met, output the best sparrow position to obtain the cluster center matrix.

[0019] Obtaining initial cluster centers according to the cluster center matrix, using Euclidean distance as a metric function, assigning patent data in the sub-dataset to the initial cluster center closest to the data, and updating cluster centers in different clusters after all patent data are assigned;

[0020] The final clustering result of the sub-dataset is obtained through iterative clustering, and the central representation of the sub-dataset is generated according to the different clusters divided.

[0021] In this solution, the patent information to be predicted is intercepted from the patent text input by the user, and a text representation of the patent information to be predicted is generated, specifically:

[0022] According to the preset paragraph position and indicated keywords, the patent text input by the user is intercepted to generate the patent information to be predicted, and the embedding vector of the patent text corresponding to the patent information to be predicted is extracted and divided into word embedding vector, sentence embedding vector and paragraph embedding vector;

[0023] The embedded vector of the patent information to be predicted is imported into the bidirectional long short-term memory network, and the attention mechanism is introduced to calculate the embedded vectors of different levels using forward LSTM and reverse LSTM. The forward and reverse calculation results are calculated through the hidden layer to output the semantic features corresponding to the embedded vector of the patent information to be predicted;

[0024] The semantic features are represented and matched with embedding vectors of different levels corresponding to the patent information to be predicted according to the position encoding to generate a text representation of the patent information to be predicted.

[0025] In this solution, constraint information is generated based on the similarity combined with patent quality indicators, specifically:

[0026] Calculate the text representation of the patent information to be predicted and all the central representations of the sub-datasets and their similarities, calculate the cosine similarity between the embedded vectors after dimensional alignment, and when the cosine similarity is greater than a preset threshold, extract the semantic features of the patent information to be predicted at the corresponding position, and use the semantic features to correct the similarity;

[0027] Traversing the patent information to be predicted to obtain all similarities, calculating the average of the absolute values ​​to generate an average similarity, taking the inverse of the average similarity to generate one of the constraint information;

[0028] Using a big data engine to obtain patent quality evaluation examples, extracting patent quality evaluation indicators from the patent quality evaluation examples, and performing principal component analysis on the patent quality evaluation indicators to identify key influencing factors;

[0029] According to the patent quality evaluation examples, the interactive relationship between key influencing factors and patent texts and between different key influencing factors is obtained, and triplets are formed based on the attributes corresponding to different interactive relationships and key influencing factors. The knowledge graph convolutional neural network is used to learn the graph structure and construct the knowledge graph.

[0030] In the knowledge graph, the number of relationship edges directly connected to the key influencing factors is obtained to calculate the centrality of the node, and the centrality is used to characterize the importance of the key influencing factors. A preset number of key influencing factors are selected according to the importance, and the corresponding indicator variable composition constraint information is obtained.

[0031] In this solution, constraint information is used to train the patent quality assessment model, specifically:

[0032] Construct a patent quality assessment model, train the corresponding encoders through the training data of different categories of patent quality indicators in the constraint information, and use the encoders of different patent quality indicators to extract indicator features from the text representation of the patent information to be predicted;

[0033] The index features of the patent information to be predicted are combined with the inverse of the average similarity between the text representation and the center representation and input into different multi-layer perceptrons to obtain the feature importance matrix, and the attention distribution of the feature importance distribution is obtained by using collaborative attention. The representation of the patent information to be predicted under different constraints is obtained according to weighted calculation;

[0034] The indicator features are fully connected with the weighted representation, and the output vector is converted into a probability distribution through the interaction of the multi-layer perceptron to obtain the prediction evaluation. The MSE evaluation indicator is used for scoring to obtain the quality assessment result of the patent information to be predicted.

[0035] The second aspect of the present invention also provides a pre-application patent quality assessment system based on cluster center representation, the system comprising: a memory, a processor, a user interaction module, an assessment data set generation module, a quality assessment module, and a data storage management module, wherein the memory and the processor store and execute a pre-application patent quality assessment method program based on cluster center representation;

[0036] The user interaction module is used for the user to input a keyword group to determine the patent data subset to be evaluated; and to input information of the patent information to be predicted as an evaluation input window; and to return the results of the system evaluation and display the evaluation results to the user;

[0037] The evaluation data set generation module generates sub-data sets based on the patent big data set according to the keyword groups provided by the user;

[0038] The quality assessment module is responsible for quality assessment based on the patent information and sub-datasets to be assessed;

[0039] The data storage management module is responsible for the storage of large patent data sets and the storage of patent subsets generated based on user keyword groups, which facilitates the operation of non-real-time evaluation tasks.

[0040] The present invention discloses a patent quality assessment method and system before application based on cluster center representation, including: extracting keywords based on patent text input by users for retrieval, generating sub-datasets with similar features in patent big data, and generating center representation of the sub-datasets through clustering models; intercepting patent information to be predicted in the patent text input by users to generate text representation; calculating the similarity between the text representation and the center representation of the patent information to be predicted, and generating constraint information based on the similarity combined with patent quality indicators; using constraint information to train patent quality assessment models, and obtaining multi-dimensional quality evaluation results for patents input by users. While solving the problem of massive data comparison, the present invention quickly conducts multi-dimensional quality analysis on patents that users plan to apply for, which is conducive to improving the success rate of user applications and cultivating high-value patents, and reducing the cost of patent applications for enterprises. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 A flow chart of a method for evaluating patent quality before application based on cluster center representation according to the present invention is shown;

[0042] Figure 2 A flow chart showing the central representation of a sub-dataset generated by the present invention is shown;

[0043] Figure 3 A flow chart showing the patent quality assessment model constructed by the present invention is shown;

[0044] Figure 4 A block diagram of the pre-application patent quality assessment system based on cluster center representation of the present invention is shown. DETAILED DESCRIPTION

[0045] In order to more clearly understand the above-mentioned purpose, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.

[0046] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the protection scope of the present invention is not limited to the specific embodiments disclosed below.

[0047] Figure 1 A flow chart of the method for evaluating patent quality before application based on cluster center representation of the present invention is shown.

[0048] like Figure 1As shown, the first aspect of the present invention provides a method for evaluating patent quality before application based on cluster center representation, comprising:

[0049] S102, extracting keywords from the patent text input by the user for retrieval, generating a sub-dataset whose feature similarity meets a preset standard in the patent big data, and generating a central representation of the sub-dataset through a clustering model;

[0050] S104, extracting the patent information to be predicted from the patent text input by the user, and generating a text representation of the patent information to be predicted;

[0051] S106, calculating the similarity between the patent information to be predicted and the central representation, and generating constraint information based on the similarity combined with the patent quality index;

[0052] S108, using the constraint information to train a patent quality assessment model, and obtaining a multi-dimensional quality evaluation result for the patent input by the user through the patent quality assessment model.

[0053] It should be noted that the patent text input by the user is obtained and preprocessed by word segmentation, normalization, and stop word screening to generate a serialized representation of the patent text, and the part-of-speech tags of the word vectors in the serialized representation of the patent text are determined, including: prepositions, adjectives, nouns, proper nouns, etc., and the part-of-speech tags are used for sequence annotation; Roberta is used to trim the serialized representation of the patent text into blocks and embed the representation. After Roberta encoding, the embedded vectors are correlated with each other, and the semantic learning ability is enhanced, the context is better connected, and the semantic features expressed in different contexts are obtained. The embedded vector of the patent text is obtained, preset phrases are filtered through the part-of-speech tags, the corresponding embedded vectors are filtered based on the position feature matching of the preset phrases, the matched and filtered embedded vectors are spliced, and after the length is unified, the dimension is transformed through the linear layer to obtain the spliced ​​embedded vector; the self-attention mechanism is introduced into the embedded vector of the patent text, the characteristics of the embedded vector are weighted by the self-attention weight, and cross-attention is introduced between the spliced ​​embedded vector and the embedded vector, the neighborhood embedded vector of the spliced ​​embedded vector is obtained, and the context semantics are enhanced; the embedded vector sequence and the neighborhood embedded vector sequence after attention encoding are obtained, wherein the embedded vector sequence contains global semantics, and the neighborhood embedded vector sequence is rich in local context semantics, the similarity of the embedded vector and the neighborhood embedded vector in the sequence after dimensionality reduction is calculated, the similarity is used as the basis for importance, and the spliced ​​embedded vector whose similarity meets the preset similarity threshold is obtained for decoding as the keyword extraction result; a search index is established according to the keyword, and the keyword feature similarity is calculated in the massive patent big data by using the search index, and the patent data that meets the preset similarity standard is obtained to construct a sub-dataset containing keywords.

[0054] Figure 2 A flow chart of the present invention for generating a central representation of a sub-dataset is shown.

[0055] According to an embodiment of the present invention, the central representation of the sub-dataset is generated by a clustering model, specifically:

[0056] S202, using the sparrow search algorithm to optimize the initial cluster center of the sub-data set, initialize the parameters of the sparrow search algorithm, calculate the fitness value in the sparrow population, and obtain the optimal fitness value and the worst fitness value and the corresponding position;

[0057] S204, selecting a discoverer, a joiner and a scout and updating their positions, introducing an adaptive t-distribution variation in the process of updating the sparrow's position, iteratively calculating the fitness and updating the sparrow's position, and outputting the best sparrow position to obtain the cluster center matrix after the maximum number of iterations is met;

[0058] S206, obtaining initial cluster centers according to the cluster center matrix, using Euclidean distance as a metric function, assigning patent data in the sub-dataset to the initial cluster center closest to the data, and updating cluster centers in different clusters after all patent data are assigned;

[0059] S208, obtaining the final clustering result of the sub-dataset through iterative clustering, and generating the central representation of the sub-dataset according to the different divided clusters.

[0060] It should be noted that the K-means clustering algorithm can improve the running speed, but when the cluster center is randomly selected, the center point is too concentrated or dispersed, which will lead to poor clustering effect and affect the accuracy of the center representation after clustering. The sparrow search algorithm has the advantages of fast convergence speed and improves the influence of the initial cluster center on the clustering result. Initialize the parameters of the sparrow search algorithm, set the maximum number of iterations, population size, number of discoverers, number of alerts and alarm value. The discoverers are composed of the sparrows with the best position, and the remaining sparrows are followers. The alerts will be randomly generated. The higher the fitness of the sparrow, the higher the priority of the sparrow to obtain food. Through the sparrow search algorithm and the input number of clusters, the cluster center matrix with the best fitness is found. In order to prevent the clustering algorithm from falling into the local optimum, the sparrow position is adaptively mutated by t distribution. The t distribution combines the characteristics of Cauchy distribution and Gaussian distribution, balances the global exploration ability and local development ability, and obtains the current latest position. If the current latest position is better than the previous optimal position, the position matrix is ​​updated until the best position matrix is ​​output to obtain the cluster center matrix. Often, there is more than one central representation after clustering. For example, after clustering analysis, N categories are obtained, corresponding to N central representations.

[0061] It should be noted that, according to the preset paragraph positions and indicative keywords such as the claims or the abstract of the specification, the patent text input by the user is intercepted to generate the patent information to be predicted, and the embedding vector of the patent text corresponding to the patent information to be predicted is extracted and generated, and the word embedding vector, sentence embedding vector and paragraph embedding vector are divided. The division of embedding vectors at different positions and levels can improve the efficiency of text semantic recognition, and the embedding vector of the patent information to be predicted is imported into a bidirectional long short-term memory network, and the attention mechanism is introduced to calculate the embedding vectors of different levels using forward LSTM and reverse LSTM. The forward LSTM performs a forward operation on the embedding vector input at time t and the output at time t-1 to obtain the forward output at time t, and the reverse LSTM performs a reverse operation on the embedding vector input at time t and the output at time t+1 to obtain the reverse output at time t. The forward and reverse calculation results are calculated through the hidden layer, and the semantic features corresponding to the embedding vector of the patent information to be predicted are output; the semantic features are represented and matched with the embedding vectors of different levels corresponding to the patent information to be predicted according to the position encoding to generate the text representation of the patent information to be predicted.

[0062] Figure 3 A flow chart of constructing a patent quality assessment model according to the present invention is shown.

[0063] According to an embodiment of the present invention, the patent quality assessment model is trained using constraint information, specifically:

[0064] S302, constructing a patent quality assessment model, training corresponding encoders with training data of different categories of patent quality indicators in the constraint information, and using encoders of different patent quality indicators to extract indicator features from the text representation of the patent information to be predicted;

[0065] S304, inputting the index features of the patent information to be predicted into different multi-layer perceptrons in combination with the inverse of the average similarity between the text representation and the center representation, obtaining a feature importance matrix, using collaborative attention to obtain the attention distribution of the feature importance distribution, and obtaining the representation of the patent information to be predicted under different constraints according to weighted calculation;

[0066] S306, fully connect the indicator features with the weighted representation, output vectors through multi-layer perceptron interaction, convert the output vectors into probability distribution to obtain prediction evaluation, score using MSE evaluation index, and obtain the quality evaluation result of the patent information to be predicted.

[0067] It should be noted that the text representation of the patent information to be predicted and all the central representations and similarities of the sub-datasets are calculated, and the cosine similarity between the embedded vectors is calculated after the dimension alignment. When the cosine similarity is greater than the preset threshold, the semantic features of the patent information to be predicted at the corresponding position are extracted, and the similarity is corrected using the semantic features; the patent information to be predicted is traversed to obtain all similarities, the mean of the absolute values ​​is calculated, and the average similarity is generated. The inverse of the average similarity can reflect the probability of authorization to a certain extent, and one of the quality evaluation indicators in the constraint information is generated; the big data engine is used to obtain patent quality evaluation examples, and patent quality evaluation indicators are extracted from the patent quality evaluation examples, such as the number and length of claims, the number of patent citations, and the number of citations of non-patent literature. The patent quality evaluation index is analyzed by principal component analysis to identify key influencing factors using quantity, technology life cycle, patent category, number of patent families, number of inventors and applicants, etc.; the interactive relationship between key influencing factors and patent texts and between different key influencing factors is obtained based on patent quality evaluation examples, and triples are formed based on different interactive relationships and attributes corresponding to key influencing factors. The knowledge graph convolutional neural network is used to learn the graph structure to build a knowledge graph; according to the characteristics of the knowledge graph, the more nodes a node is connected to, the richer the implicit information of the node may be. The number of relationship edges directly connected to the key influencing factors in the knowledge graph is obtained to calculate the centrality of the node, and the centrality is used to characterize the importance of the key influencing factors. The more central nodes will be more important than other nodes. A preset number of key influencing factors are selected according to the importance, and the corresponding indicator variable composition constraint information is obtained.

[0068] A patent quality assessment model is constructed. The multi-scale encoder module is used to extract index features from the text representation of the patent information to be predicted, and the representation of the text representation under different patent evaluation index variables in the constraint information is obtained. The collaborative attention mechanism is used to estimate the different importance of different representations. The calculation formula is: ,in represents the attention distribution, It represents the representation of the text representation corresponding to the jth and nth patent evaluation indicator variables, and m represents the total number. The indicator features before the attention mechanism are connected to the indicator representation after the attention mechanism respectively, and the results are output through the multi-layer perceptron interaction network. The MSE evaluation index is used for scoring to obtain the quality evaluation results of the patent information to be predicted. The smaller the MSE evaluation index is, the closer the predicted value output by the quality evaluation model is to the true value, which proves that the quality of the patent information to be predicted is better.

[0069] It should be noted that the historical quality assessment results of the enterprise's patent texts are obtained, and a writing portrait of the enterprise's patent texts is constructed based on the historical quality assessment results. A personalized database is constructed based on the writing portraits of different time periods. The quality evaluation indicators in the writing portraits that are significantly different from other quality evaluation indicators are obtained and marked in the personalized database. According to the marked quality evaluation indicators, improvement directions for patent text writing are generated in the current writing workflow. The improvement directions are traced based on the ant colony algorithm using the improvement directions, and the influencing factors of abnormal quality evaluation indicators are obtained according to the tracing path. According to the influencing factors, optimization measures are retrieved to improve the writing process of patent texts and technical briefing documents, and the corresponding writing workflow is updated.

[0070] Figure 4 A block diagram of the pre-application patent quality assessment system based on cluster center representation of the present invention is shown.

[0071] The second aspect of the present invention also provides a pre-application patent quality assessment system 4 based on cluster center representation, the system comprising: a memory 41, a processor 42, a user interaction module 43, an assessment data set generation module 44, a quality assessment module 45, and a data storage management module 46, wherein the memory 41 and the processor 42 store and execute a pre-application patent quality assessment method program based on cluster center representation;

[0072] The user interaction module 43 is used for the user to input a keyword group to determine the patent data subset to be evaluated; and to input information of the patent information to be predicted as an evaluation input window; and to return the result of the system evaluation and display the evaluation result to the user;

[0073] An evaluation data set generation module 44 generates a sub-data set based on the patent big data set according to the keyword group provided by the user;

[0074] A quality assessment module 45, responsible for performing quality assessment based on the patent information to be assessed and the sub-datasets;

[0075] The data storage management module 46 is responsible for the storage of large patent data sets and the storage of patent subsets generated based on user keyword groups, so as to facilitate the operation of non-real-time evaluation tasks.

[0076] The third aspect of the present invention also provides a computer-readable storage medium, which includes a pre-application patent quality assessment method program based on cluster center representation. When the pre-application patent quality assessment method program based on cluster center representation is executed by a processor, the steps of the pre-application patent quality assessment method based on cluster center representation as described in any one of the above items are implemented.

[0077] In the several embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as: multiple units or components can be combined, or can be integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the components shown or discussed can be through some interfaces, and the indirect coupling or communication connection of the devices or units can be electrical, mechanical or other forms.

[0078] The units described above as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units; they may be located in one place or distributed on multiple network units; some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0079] In addition, all functional units in the embodiments of the present invention may be integrated into one processing unit, or each unit may be separately used as a unit, or two or more units may be integrated into one unit; the above-mentioned integrated units may be implemented in the form of hardware or in the form of hardware plus software functional units.

[0080] A person of ordinary skill in the art can understand that: all or part of the steps of implementing the above method embodiment can be completed by hardware related to program instructions, and the aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it executes the steps of the above method embodiment; and the aforementioned storage medium includes: a mobile storage device, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and other media that can store program codes.

[0081] Alternatively, if the above-mentioned integrated unit of the present invention is implemented in the form of a software function module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiment of the present invention can be essentially or partly reflected in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the methods described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as mobile storage devices, ROM, RAM, magnetic disks or optical disks.

[0082] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A pre-application patent quality assessment method based on cluster center representation, characterized in that Extract keywords from patent texts input by users for retrieval, generate sub-datasets whose feature similarity meets preset standards in patent big data, and generate central representations of the sub-datasets through clustering models; Extracting the patent information to be predicted from the patent text input by the user, and generating a text representation of the patent information to be predicted; Calculate the similarity between the patent information to be predicted and the central representation, and generate constraint information based on the similarity combined with the patent quality index; Using constraint information to train a patent quality assessment model, and obtaining a multi-dimensional quality evaluation result for a patent input by a user through the patent quality assessment model; Based on the similarity and patent quality indicators, constraint information is generated, specifically: Calculate the similarity between the text representation of the patent information to be predicted and the central representation of all sub-datasets, calculate the cosine similarity between the embedded vectors after dimensional alignment, and when the cosine similarity is greater than a preset threshold, extract the semantic features of the patent information to be predicted at the corresponding position, and use the semantic features to correct the similarity; Traversing the patent information to be predicted to obtain all similarities, calculating the average of the absolute values ​​to generate an average similarity, taking the inverse of the average similarity to generate one of the constraint information; Using a big data engine to obtain patent quality evaluation examples, extracting patent quality evaluation indicators from the patent quality evaluation examples, and performing principal component analysis on the patent quality evaluation indicators to identify key influencing factors; According to the patent quality evaluation examples, the interactive relationship between key influencing factors and patent texts and between different key influencing factors is obtained, and triplets are formed based on the attributes corresponding to different interactive relationships and key influencing factors. The knowledge graph convolutional neural network is used to learn the graph structure and construct the knowledge graph. Obtain the number of relationship edges directly connected to the key influencing factors in the knowledge graph to calculate the centrality of the node, use the centrality to characterize the importance of the key influencing factors, select a preset number of key influencing factors according to the importance, and obtain the corresponding indicator variable composition constraint information; The patent quality assessment model is trained using constraint information, specifically: Construct a patent quality assessment model, train the corresponding encoders with the training data of different categories of patent quality indicators in the constraint information, and use the encoders of different patent quality indicators to extract indicator features from the text representation of the patent information to be predicted; The index features of the patent information to be predicted are combined with the inverse of the average similarity between the text representation and the center representation and input into different multi-layer perceptrons to obtain the feature importance matrix, and the attention distribution of the feature importance distribution is obtained by using collaborative attention. The representation of the patent information to be predicted under different constraints is obtained according to weighted calculation; The indicator features are fully connected with the weighted representation, and the output vector is converted into a probability distribution through the interaction of the multi-layer perceptron to obtain the prediction evaluation. The MSE evaluation index is used for scoring to obtain the quality assessment result of the patent information to be predicted.

2. The method for evaluating patent quality before application based on cluster center representation according to claim 1, characterized in that: Based on the patent text input by the user, keywords are extracted for retrieval, and sub-datasets with similar features that meet the preset standards are generated in the patent big data, specifically: Obtain the patent text input by the user for word segmentation preprocessing, generate a serialized representation of the patent text, determine the part-of-speech tags of the word vectors in the serialized representation of the patent text, and use the part-of-speech tags for sequence annotation; Using Roberta to trim, segment and embed the serialized representation of the patent text, obtain an embedding vector of the patent text, filter preset phrases through the part-of-speech tags, filter corresponding embedding vectors based on position feature matching of the preset phrases, and splice the matched and filtered embedding vectors to obtain a spliced ​​embedding vector; A self-attention mechanism is introduced into the embedding vector of the patent text. The features of the embedding vector are strengthened by weighting the self-attention weight. Cross-attention is introduced between the concatenated embedding vector and the embedded vector to obtain the neighborhood embedding vector of the concatenated embedding vector and strengthen the contextual semantics. Obtain the embedded vector sequence and the neighborhood embedded vector sequence after attention encoding, calculate the similarity of the embedded vector and the neighborhood embedded vector in the sequence, obtain the concatenated embedded vector whose similarity meets the preset similarity threshold for decoding, and use it as the keyword extraction result; A search index is established based on the keywords, and the keyword feature similarity is calculated in the massive patent big data using the search index to obtain patent data that meets the preset similarity standard to construct a sub-dataset containing the keywords.

3. The method for evaluating patent quality before application based on cluster center representation according to claim 1, characterized in that: The central representation of the sub-dataset is generated by the clustering model, specifically: Use the sparrow search algorithm to optimize the initial cluster center of the sub-dataset, initialize the parameters of the sparrow search algorithm, calculate the fitness value in the sparrow population, and obtain the optimal fitness value and the worst fitness value and the corresponding position; Select the discoverer, joiner and scout and update the position. In the process of updating the sparrow's position, introduce the adaptive t-distribution variation, iteratively calculate the fitness and update the sparrow's position. After the maximum number of iterations is met, output the best sparrow position to obtain the cluster center matrix. Obtaining initial cluster centers according to the cluster center matrix, using Euclidean distance as a metric function, assigning patent data in the sub-dataset to the initial cluster center closest to the data, and updating cluster centers in different clusters after all patent data are assigned; The final clustering result of the sub-dataset is obtained through iterative clustering, and the central representation of the sub-dataset is generated according to the different clusters divided.

4. The method for evaluating patent quality before application based on cluster center representation according to claim 1, characterized in that: The patent information to be predicted is intercepted from the patent text input by the user, and a text representation of the patent information to be predicted is generated, specifically: According to the preset paragraph position and indicated keywords, the patent text input by the user is intercepted to generate the patent information to be predicted, and the embedding vector of the patent text corresponding to the patent information to be predicted is extracted and divided into word embedding vector, sentence embedding vector and paragraph embedding vector; The embedded vector of the patent information to be predicted is imported into the bidirectional long short-term memory network, and the attention mechanism is introduced to calculate the embedded vectors of different levels using forward LSTM and reverse LSTM. The forward and reverse calculation results are calculated through the hidden layer to output the semantic features corresponding to the embedded vector of the patent information to be predicted; The semantic features are represented and matched with embedding vectors of different levels corresponding to the patent information to be predicted according to the position encoding to generate a text representation of the patent information to be predicted.

5. A pre-application patent quality assessment system based on cluster center representation, characterized in that: Implementing the method for evaluating patent quality before application based on cluster center representation as described in any one of claims 1 to 4, the system comprises: a memory, a processor, a user interaction module, an evaluation data set generation module, a quality evaluation module, and a data storage management module, wherein the program for evaluating patent quality before application based on cluster center representation is stored and executed in the memory and the processor; The user interaction module is used for the user to input a keyword group to determine the patent data subset to be evaluated; and to input information of the patent information to be predicted as an evaluation input window; and to return the results of the system evaluation and display the evaluation results to the user; The evaluation data set generation module generates sub-data sets based on the patent big data set according to the keyword groups provided by the user; The quality assessment module is responsible for quality assessment based on the patent information and sub-datasets to be assessed; The data storage management module is responsible for the storage of large patent data sets and the storage of patent subsets generated based on user keyword groups, which facilitates the operation of non-real-time evaluation tasks.

6. A pre-application patent quality assessment system based on cluster center representation according to claim 5, characterized in that: The central representation of the sub-dataset is generated in the evaluation dataset generation module, specifically: Use the sparrow search algorithm to optimize the initial cluster center of the sub-dataset, initialize the parameters of the sparrow search algorithm, calculate the fitness value in the sparrow population, and obtain the optimal fitness value and the worst fitness value and the corresponding position; Select the discoverer, joiner and scout and update the position. In the process of updating the sparrow's position, introduce the adaptive t-distribution variation, iteratively calculate the fitness and update the sparrow's position. After the maximum number of iterations is met, output the best sparrow position to obtain the cluster center matrix. Obtaining initial cluster centers according to the cluster center matrix, using Euclidean distance as a metric function, assigning patent data in the sub-dataset to the initial cluster center closest to the data, and updating cluster centers in different clusters after all patent data are assigned; The final clustering result of the sub-dataset is obtained through iterative clustering, and the central representation of the sub-dataset is generated according to the different clusters divided.

7. The pre-application patent quality assessment system based on cluster center representation according to claim 5, characterized in that: Obtain the constraint information of the patent quality assessment model in the quality assessment module, specifically: Calculate the text representation of the patent information to be predicted and all the central representations of the sub-datasets and their similarities, calculate the cosine similarity between the embedded vectors after dimensional alignment, and when the cosine similarity is greater than a preset threshold, extract the semantic features of the patent information to be predicted at the corresponding position, and use the semantic features to correct the similarity; Traversing the patent information to be predicted to obtain all similarities, calculating the average of the absolute values ​​to generate an average similarity, taking the inverse of the average similarity to generate one of the constraint information; Using a big data engine to obtain patent quality evaluation examples, extracting patent quality evaluation indicators from the patent quality evaluation examples, and performing principal component analysis on the patent quality evaluation indicators to identify key influencing factors; According to the patent quality evaluation examples, the interactive relationship between key influencing factors and patent texts and between different key influencing factors is obtained, and triplets are formed based on the attributes corresponding to different interactive relationships and key influencing factors. The knowledge graph convolutional neural network is used to learn the graph structure and construct the knowledge graph. In the knowledge graph, the number of relationship edges directly connected to the key influencing factors is obtained to calculate the centrality of the node, and the centrality is used to characterize the importance of the key influencing factors. A preset number of key influencing factors are selected according to the importance, and the corresponding indicator variable composition constraint information is obtained.

8. The pre-application patent quality assessment system based on cluster center representation according to claim 5, characterized in that: The patent quality assessment model in the quality assessment module is specifically: Construct a patent quality assessment model, train the corresponding encoders with the training data of different categories of patent quality indicators in the constraint information, and use the encoders of different patent quality indicators to extract indicator features from the text representation of the patent information to be predicted; The index features of the patent information to be predicted are combined with the inverse of the average similarity between the text representation and the center representation and input into different multi-layer perceptrons to obtain the feature importance matrix, and the attention distribution of the feature importance distribution is obtained by using collaborative attention. The representation of the patent information to be predicted under different constraints is obtained according to weighted calculation; The indicator features are fully connected with the weighted representation, and the output vector is converted into a probability distribution through the interaction of the multi-layer perceptron to obtain the prediction evaluation. The MSE evaluation index is used for scoring to obtain the quality assessment result of the patent information to be predicted.

Images