A patent cross-domain recommendation method and system fusing Doc2vec and GAT

Abstract

The application discloses a patent cross-field recommendation method and system fusing Doc2vec and GAT, and the method comprises the following steps: S100, data acquisition; collecting the publication number, title, abstract and reference relationship of a patent from an open source patent platform, and pre-processing the text content; S200, model training; learning the patent content by using Doc2vec, and training a GAT model to fuse the patent content and the reference relationship; S300, retrieval and recommendation; inputting a design requirement into the trained model, vectorizing the design requirement, and obtaining a pushed cross-field patent. The technology of the application can extract the function, technical terms and field information of the patent, and can also learn how the patent inventor learns from the information of the existing patent, so that the effectiveness of the patent cross-field recommendation is improved.

Description

Technical Field

[0001] This invention relates to the fields of computer network and big data processing technology, and in particular to a patented cross-domain recommendation method and system that integrates Doc2vec and GAT for innovative design of complex products. Background Technology

[0002] Complex industrial products such as airplanes, trains, and ships are typically characterized by a wide variety of products, small batch sizes, complex system components, and high development costs. Their innovative design is a knowledge-intensive activity, often involving cross-disciplinary knowledge / technology transfer and application. However, while designers of complex industrial products each possess unique strengths, it's difficult for them to simultaneously master all the knowledge required for the entire design process. Patents cover 90% to 95% of the world's technical information and are a crucial source for expanding design knowledge and innovating products. Therefore, designers urgently need cross-disciplinary patent recommendations during the design process of complex industrial products.

[0003] Cross-domain patent recommendation tasks primarily address situations where different patents have similar inventive functions but different application areas, requiring analysis of patent text content and patent citation relationships. Patent text analysis aims to identify the function and field of the patent, while patent cross-citation relationship analysis aims to uncover whether existing patents can be recommended across domains. The article "A new function-based patent knowledge retrieval tool for conceptual design of innovative products" in Volume 115 of the journal *Computers in Industry* provides a patent retrieval tool based on patent function, technical terminology, and IPC tags. This method uses a semi-supervised algorithm to annotate functional information in patents and sets three evaluation indicators—maturity, clarity, and scalability—to assess whether a patent can be recommended across domains.

[0004] However, the article "A new function-based patent knowledge retrieval tool for conceptual design of innovative products" in Volume 115 of the journal Computers in Industry provides a patent retrieval tool based on patent function, technical terminology, and IPC tags. This tool only analyzes the patent text and ignores the implicit cross-domain knowledge information contained in the patent citation information (i.e., there are patents that use the same technology but are used in different fields). Summary of the Invention

[0005] To address the problems existing in current technologies, the purpose of this invention is to propose a cross-domain knowledge recommendation method that comprehensively analyzes patent text content and cross-citation relationships. This invention aims to train an algorithmic model that integrates Doc2vec and GAT, enabling it to extract not only the functions, technical terms, and domain information of patents, but also to learn how patent inventors reference information from existing patents, thereby improving the effectiveness of cross-domain patent recommendations.

[0006] To achieve the above objectives, this invention provides a patent cross-domain recommendation method integrating Doc2vec and GAT, comprising the following steps:

[0007] Step S100: Data collection; Collect the publication number, title, abstract and citation relationship of the patent from the open source patent platform, and preprocess the text content;

[0008] Step S200: Model training; Use Doc2vec to learn the patent content and train the GAT model to fuse the patent content and citation relationships;

[0009] Step S300: Search and recommend; input the design requirements into the trained model, vectorize the design requirements, and obtain the recommended cross-domain patents.

[0010] Furthermore, step S200 includes:

[0011] Step S2: Use the BERT model to perform named entity recognition on the patent abstract;

[0012] Step S3: Determine whether the collected patent abstracts indicate the field, function, or technical information.

[0013] Step S4: If the patent does not contain any information in the field, function, or technology, it means that the patent is not related to the innovative design of complex products, and the patent will not be trained for the time being.

[0014] Step S5: If the patent is marked with field, functional information, or technical terms, then use Doc2vec to learn the patent content;

[0015] Step S6: Construct a patent cross-citation network based on patent citation relationships;

[0016] Step S7: Embed the patent citation network information with the patent's field, function, and technical terms into a vector VEC. D ,VEC F ,VEC T Three GAT models are input for parameter tuning and training to update the patent embedding vectors;

[0017] Step S8: [The text appears to be incomplete and contains several typographical errors. A more accurate translation would require thed ,vec' f ,vec' t By splicing them together, we get <vec' d ,vec' f ,vec' t A vector is an embedded vector that contains domain information, functional information, technical terminology information, and reference information.

[0018] Furthermore, in step S2, firstly, high-quality named entity recognition data in a certain field is manually annotated, and the field, functional information, and technical terms in the patent abstract are labeled as D, F, and T, respectively; where D represents the field, F represents the function, and T represents the technical term. The manually annotated named entity recognition dataset is then input into the BERT model for parameter tuning and learning. The main parameters to be adjusted are sequence length, batch size, and number of iterations. After that, the trained BERT model is used to perform named entity recognition on the unannotated patent abstracts, and annotate the field, functional information, or technical term information.

[0019] Furthermore, in step S5, the specific method for learning the patent content using Doc2vec is as follows: merging the text content of the patent title and abstract to train three Doc2vec models, and training only one Doc2vec model using the patent title, for a total of four Doc2vec models. Simultaneously, each patent also obtains three content embedding vectors (vec). d ,vec f ,vec t This forms a content embedding vector set VEC. D ,VEC F ,VEC T and a title embedding vector vec b The corresponding title embedding vector set VEC B The main parameters adjusted during the training of the Doc2vec model are the sliding window size, learning rate, and number of iterations. After each training session, the model is evaluated based on five pre-specified metrics: accuracy, recall, F1 score, normalized loss cumulative gain, and mean precision, to determine the optimal model.

[0020] Furthermore, in step S6, a patent citation network is constructed based on patent citation relationships; if there is a citation between two patents, then the two are connected by an edge in the network, forming a patent citation network of <initial patent, cite, first-level citation patent, cite, second-level citation patent>.

[0021] Furthermore, in step S7, the patent embedding vector is updated so that the vector vec of each patent is updated. d ,vec f ,vec tThe reference relationship information is incorporated and represented as vec' d ,vec' f ,vec' t This refers to the embedding vectors of <domain information + reference information>, <functional information + reference information>, and <technical terminology information + reference information>.

[0022] Furthermore, step S300 includes:

[0023] Step S9: The user inputs the design requirements to be addressed;

[0024] Step S10: Vectorize the design requirements; use the Doc2vec model to transform the user-input design requirements into a design requirement vector, represented as vec. r , r represents demand;

[0025] Step S11: Identify the patents that represent the design requirements;

[0026] Step S12: Retrieve patents that are technically or functionally related but in different fields from the trained patent embedding vector model;

[0027] Step S13: Push the patents ranked first in the "Design Requirements - Patents" similarity list to the user.

[0028] Furthermore, in step S11, vec r respectively in vec b Similarity calculations are performed, and the five patents with the highest similarity are pushed to the user. After reading them, the user selects one patent as an alternative for their design requirements. The vector representation of this patent is vec. k .

[0029] Furthermore, in step S12, the specific operation method is as follows: retrieve the results related to vec. k vec is calculated separately for patents in different fields. k The similarity between these patent vectors is calculated and sorted from highest to lowest similarity to obtain a "Design Requirement - Patent" similarity list.

[0030] On the other hand, the present invention provides a patent cross-domain recommendation system integrating Doc2vec and GAT, the system being used to implement the method according to any one of claims 1-9, the system comprising:

[0031] The data acquisition module is used to acquire patent information, which includes the patent publication number, title, abstract, and citation relationships.

[0032] Model training module; uses Doc2vec to learn from patent content and trains the GAT model to fuse patent content and citation relationships;

[0033] The retrieval and recommendation module inputs design requirements into the trained model, vectorizes the design requirements, and obtains cross-domain patent recommendations.

[0034] The beneficial effects of this invention are as follows: Compared with existing patent retrieval and recommendation technologies, the technical solution proposed in this invention comprehensively considers the functional, domain, and technical terminology features and patent citation relationships within the patent content, resulting in more comprehensive information learned by the model. Simultaneously, the model performs better. Experimental verification with a Doc2vec model that only learns patent text content and a scheme combining Doc2vec and a Graph Convolutional Network (GCN) model without attention mechanisms reveals that the patent recommendation scheme combining Doc2vec and GAT proposed in this invention outperforms the model in terms of precision, recall, F1 score, Normalized Discounted Cumulative Gain (NDCG), and mean average precision (mAP). Compared with existing patent retrieval and recommendation technologies, the technical solution proposed in this invention can handle design tasks based on natural language, eliminating the need for designers / engineers to manually input patent keywords for retrieval, making patent retrieval and recommendation easier. Furthermore, this technical solution enables designers to refer to high-tech patents in other fields when conducting innovative designs, achieving cross-domain knowledge recommendation. Attached Figure Description

[0035] Figure 1 A flowchart of a patent cross-domain recommendation method integrating Doc2vec and GAT according to an embodiment of the present invention is shown;

[0036] Figure 2 A schematic diagram of the structure of the patented embedding vector learning model according to an embodiment of the present invention is shown;

[0037] Figure 3 A schematic diagram of the cross-domain recommendation system structure according to an embodiment of the present invention is shown. Detailed Implementation

[0038] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0039] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0040] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0041] The following combination Figures 1-3 Specific embodiments of the present invention will be described in detail below. It should be understood that the specific embodiments described herein are for illustrative and explanatory purposes only and are not intended to limit the present invention.

[0042] This invention provides a patent cross-domain recommendation method and system that integrates Doc2vec and GAT, for cross-domain knowledge recommendation by comprehensively analyzing patent text content and mutual citation relationships.

[0043] The patent cross-domain recommendation system integrating Doc2vec and GAT according to the present invention includes:

[0044] The data acquisition module is used to acquire patent information, which includes the patent publication number, title, abstract, and citation relationships.

[0045] Model training module; uses Doc2vec to learn from patent content and trains the GAT model to fuse patent content and citation relationships;

[0046] The retrieval and recommendation module inputs design requirements into the trained model, vectorizes the design requirements, and obtains cross-domain patent recommendations.

[0047] Figure 1 This is the main flowchart of the patented cross-domain recommendation method of this invention, which integrates Doc2vec and GAT.

[0048] like Figure 1 As shown, the patent cross-domain recommendation method integrating Doc2vec and GAT proposed in this invention includes the following steps:

[0049] Step S100: Data Collection; Collect the publication number, title, abstract, and citation relationships of patents from an open-source patent platform, and preprocess the text content, including standardizing language formatting and removing irrelevant symbols. The open-source patent platform may be Google Patents.

[0050] Step S200: Model training; Use Doc2vec to learn the patent content and train the GAT model to integrate the patent content and citation relationships; specifically including the following steps S2-S8.

[0051] Step S2: Perform named entity recognition on the patent abstracts using the BERT model. BERT (Bidirectional Encoder Representations from Transformers) is a pre-trained language model based on the Transformer architecture, supporting various natural language processing tasks. To complete this step, the BERT model is tuned and trained using a named entity recognition sample dataset. The sample dataset contains partial patent abstracts, where words or phrases representing domain information, functional information, and technical information are labeled D, F, and T, respectively. D represents the domain of the technology object used in the patent, F represents the main inventive function of the patent, and T represents the technology object used in the patent. For example, in one embodiment, “An annular elastomeric EMI shielding gasket includes a metallic mesh embedded in an elastomeric sheet for sealing and protecting aircraft fuel tanks and panels.”, the field of the technology object used in the patent, “aircraft,” is labeled D; the main inventive function of the patent, “EMI shielding,” is labeled F; and the technology object used in the patent, “fuel tanks,” is labeled T. The main parameters adjusted are sequence length, batch size, and number of iterations. Then, a trained BERT model is used to perform named entity recognition on the unlabeled patent abstract, labeling the field, function, and technical information.

[0052] Step S3: Determine whether the collected patent abstracts indicate the field, function, or technical information.

[0053] Step S4: If the patent does not contain any information in the field, function, or technology, it means that the patent is not related to the innovative design of complex products, and the patent will not be trained for the time being.

[0054] Step S5: If the patent includes domain, functional information, or technical terms, then the PV-DM model of Doc2vec is used to learn the patent content. Doc2vec is an unsupervised algorithm used to map text information into fixed-length vector representations. The PV-DM model is one type of text vectorization model that predicts the center word based on context using a sliding window, such as... Figure 2 As shown, D is the paragraph vector used to mark the paragraph where the sliding window is located, W is the word vector of the context words, and C is the word vector of the predicted center word. The specific operation method is as follows: merge the text content of the patent title and abstract to train three Doc2vec models, which are used to provide the input node vectors for three GAT models related to domain (D), function (F), and technology (T), respectively. Simultaneously, train only one Doc2vec model using the patent title, for a total of four Doc2vec models. Each patent also obtains three content embedding vectors (i.e., each patent has a vec vector). d ,vec f ,vec t This forms a content embedding vector set VEC. D ,VEC F ,VEC T ) and a title embedding vector (i.e., vec b The corresponding title embedding vector set VEC B The main parameters adjusted during Doc2vec model training are sliding window size, learning rate, and number of iterations. After each training iteration, the model is evaluated based on the degree of consistency between the predicted and actual patent citation / citation lists. This evaluation is conducted using five pre-specified metrics: precision, recall, F1 score, Normalized Discounted Cumulative Gain (NDCG), and mean average precision (mAP) to determine the optimal model. The calculation methods for each metric are shown in Table 1.

[0055]

[0056] According to Table 1, the formula for calculating precision is:

[0057]

[0058] Where TP represents the number of correctly predicted patents in the prediction list (the same applies below); FP represents the number of incorrectly predicted patents in the prediction list.

[0059] The formula for calculating recall is:

[0060]

[0061] Wherein, FN represents the number of patents not predicted in the actual citation / cited list.

[0062] The formula for calculating the F1 score is:

[0063]

[0064] The formula for calculating the normalized loss cumulative gain (NDCG) is as follows:

[0065]

[0066] Where k is the predicted length of the reference list.

[0067] The formula for calculating the mean precision (mAP, which is the arithmetic mean of AP for all sampled data) is:

[0068]

[0069] Where TP@i is the TP at point i in the summation operation; n c This is the actual length of the reference list.

[0070] Step S6: Construct a patent cross-citation network based on patent citation relationships. If two patents cite each other, they are connected by an edge in the network. For example, if patent A cites patents B and C, then there is an edge connecting them.<A,cite,B> and<A,cite,C> Furthermore, if patent B cites patent D, then an edge exists.<B,cite,D> To control network size, 100 patents were initially collected manually, and their secondary citation relationships were collected to form a patent citation network of <initial patent, cite, primary citation patent, cite, secondary citation patent>.

[0071] Step S7: Embed the patent citation network information into the VEC vector along with the patent's field (D), function (F), and technical term (T). D ,VEC F ,VEC TThree GAT models are input for parameter tuning and training to update the patent embedding vector. GAT (Graph Attention Network) is a graph neural network model based on an attention mechanism that adaptively learns the importance weights between each node and its neighboring nodes, and captures more complex relationships between nodes through the attention mechanism, such as... Figure 2 As shown, "P1, P2, ..." represent patent nodes. The graph attention layer updates the vectors of each node by aggregating the feature information of neighboring nodes, and the fully connected layer maps the updated node vectors to the target output vector. The main parameters adjusted during GAT model training are the learning rate, the number of hidden units, the number of attention heads, and the number of iterations. After each training iteration, the model is evaluated based on five pre-specified metrics to determine the optimal model. Through this step, the vector vec of each patent is obtained. d ,vec f ,vec t The reference relationship information is incorporated and represented as vec' d ,vec' f ,vec' t This refers to the embedding vectors of <domain information + reference information>, <functional information + reference information>, and <technical terminology information + reference information>.

[0072] Step S8: [The text appears to be incomplete and contains several typographical errors. A more accurate translation would require the d ,vec' f ,vec' t By splicing them together, we get <vec' d ,vec' f ,vec' t A vector is an embedded vector that contains domain information, functional information, technical terminology information, and reference information.

[0073] Step S300: Search and recommend; input the design requirements into the trained model, vectorize the design requirements, and obtain the recommended cross-domain patents, specifically including the following steps S9-S13.

[0074] Step S9: Have the user input the design requirements to be addressed, which can be described in natural language.

[0075] Step S10: Vectorize the design requirements. Use the Doc2vec model to transform the user-input design requirements into a design requirement vector, represented as vec. r , r represents requirement.

[0076] Step S11: Identify the patents representing the design requirements. (The text abruptly ends here, likely due to an incomplete sentence or a formatting error.) r respectively in vec bSimilarity calculation is performed using the following formula: The five patents with the highest similarity are presented to the user. After reading them, the user selects one patent as an alternative to their design requirement. The vector representation of this patent is vec. k .

[0077] Step S12: Retrieve patents that are technically or functionally related but in different fields from the trained patent embedding vector model. Specifically, retrieve patents related to vec... k vec is calculated separately for patents in different fields. k The similarity to these patent vectors is calculated using the following formula: The similarity scores are then sorted from highest to lowest to obtain a "Design Requirements - Patents" similarity list.

[0078] Step S13: Push the patents ranked first in the "Design Requirements - Patents" similarity list to the user.

[0079] Compared with existing patent search and recommendation technologies, the technical solution proposed in this invention comprehensively considers feature information such as function, field, and technical terms in the patent content, as well as patent citation relationships, resulting in more comprehensive information learned by the model. Furthermore, the model performs better. Based on the degree of consistency between the predicted patent citation / cited list and the actual citation / cited list, comparative experiments were conducted with the Doc2vec model that only learns patent text content information and the scheme combining Doc2vec and Graph Convolutional Network (GCN) without attention mechanism. The experimental results are shown in Table 2. The first row of the horizontal axis represents five evaluation metrics: precision, recall, F1 score, Normalized Discounted Cumulative Gain (NDCG), and mean average precision (mAP). The second row of the horizontal axis represents the length of the predicted citation / cited list. The first column of the vertical axis represents the prediction models for domain, function, and technology, respectively, and the second column represents the three specific model combination schemes. All experimental data are percentages. For simplicity, GCN and GAT represent the scheme combining Doc2vec and GCN, and the scheme combining Doc2vec and GAT, respectively. The experimental results show that the patent recommendation scheme combining Doc2vec and GAT proposed in this invention performs better in all five evaluation metrics. Compared to existing patent search and recommendation technologies, the technical solution proposed in this invention can handle design task requirements based on natural language, eliminating the need for designers / engineers to manually input patent keywords for searching, thus making patent search and recommendation easier. Simultaneously, this technical solution enables designers to refer to high-tech patents in other fields during innovative design, achieving cross-domain knowledge recommendation. Furthermore, this technical solution can identify patent characteristics such as field, function, and technical terms, assisting designers / engineers in understanding patent content. Finally, this technical solution can be applied to the design of sealing structures for aircraft fuel tanks, such as... Figure 3 As shown, when designer 2 receives a design task 1 for the sealing structure of an aircraft fuel tank, designer 2 can upload the task 1 to the cross-domain knowledge recommendation system 3. The system will analyze the design task requirements, search the patent database 5, and recommend patents 6 from different fields with the same function and technology as the design task through the cross-domain patent recommendation algorithm model 5 for designer 2 to refer to. This solves the problem that when designer 2 needs to refer to high-tech patents in other fields, but does not know the technical keywords in other fields, it is difficult to find suitable patents.

[0080] Table 2 is shown below:

[0081]

[0082] Table 2

[0083] The following is another embodiment of the present invention.

[0084] Step 1: Using aircraft, vehicles, ships, and sealing as keywords, collect the publication numbers, titles, abstracts, and citation relationships of 17,653 sealing technology-related patents worldwide from the Google Patents open-source patent platform using data collection tools. After translating the patent titles and abstracts into English, remove irrelevant symbols such as drawing numbers and trademark numbers.

[0085] Step 2: Perform named entity recognition on patent abstracts using the BERT model. First, the BERT model is tuned and trained using a named entity recognition sample dataset. This dataset contains a portion of patent abstracts, approximately 10% of the total dataset. Words or phrases representing domain, function, and technology information are labeled D, F, and T, respectively. D represents the domain of the patent's technical object, F represents the patent's main inventive function, and T represents the patent's technology. The optimal parameters for training the BERT model are: sequence length 128, batch size 32, and number of iterations 30. Then, the trained BERT model is used to perform named entity recognition on unlabeled patent abstracts, labeling them with domain, function, and technical terms.

[0086] Step 3: Determine whether the collected patent abstracts include field, functional information, or technical terms.

[0087] Step 4: If the patent lacks field, functional information, and technical terminology information, it indicates that the patent is irrelevant to the target task, and the patent will not be trained at this time. In this embodiment, this section contains 5425 patents.

[0088] Step 5: If the patent is labeled with field, functional information, or technical terms, then the PV-DM model of Doc2vec is used to learn the patent content. In this embodiment, this part contains 12,228 patents. Specifically, the text content of the patent title and abstract is merged to train three Doc2vec models, and only one Doc2vec model is trained using the patent title, for a total of four Doc2vec models. Simultaneously, each patent also obtains three content embedding vectors (i.e., each patent has a vec). d ,vec f ,vec t This forms a content embedding vector set VEC. D ,VEC F ,VEC T ) and a title embedding vector (i.e., vecb The corresponding title embedding vector set VEC B The main parameters set for training the optimal Doc2vec model were a sliding window size of 15, a learning rate of 0.001, and 12000 iterations.

[0089] Step 6: Construct a patent cross-citation network based on patent citation relationships. If two patents are cited from each other, they are connected by an edge in the network. To control the network size, 100 high-quality patents in the field are initially collected manually, and the secondary citation relationships of these 100 patents are collected to form a patent citation network of <initial patent, cite, first-level citing patent, cite, second-level citing patent>.

[0090] Step 7: Embed the patent citation network information with the patent's field (D), function (F), and technical term (T) into the vector VEC. D ,VEC F ,VEC T Three GAT models were input for parameter tuning and training to update the patent embedding vectors. The main parameters of the optimal GAT model were set as follows: learning rate 0.0005, number of hidden units 128, number of attention heads 8, and number of iterations 250. Through this step, the vector vec of each patent was obtained. d ,vec f ,vec t The reference relationship information is incorporated and represented as vec' d ,vec' f ,vec' t .

[0091] Step 8: [The text appears to be incomplete and contains several typographical errors. A more accurate translation would require the full context.] d ,vec' f ,vec' t By splicing them together, we get <vec' d ,vec' f ,vec' t A vector is an embedded vector that contains domain information, functional information, technical terminology information, and reference information.

[0092] Step 9: The user inputs the design requirement to be addressed: "Aircraft composite fuel tank sealing".

[0093] Step 10: Vectorize the design requirements. Use the Doc2vec model to transform the user-input design requirements into a design requirement vector, represented as vec. r .

[0094] Step 11: Identify patents that represent the design requirements. (The text abruptly ends here, likely due to an incomplete sentence or a formatting error.) r respectively in vec bSimilarity calculation is performed using the following formula: The five patents with the highest similarity were presented to the user. After reading them, the user selected patent CN106768726A as an alternative to their design requirements. The title of this patent is "A kind of method for technique for aircraft composite fuel tank sealing propertytest." Considering the technical field of this patent, it belongs to the "aircraft" field. The vector representation of this patent is vec. k .

[0095] Step 12: Retrieve patents that are technically or functionally related but in different fields from the trained patent embedding vector model. Specifically, this involves retrieving patents related to vec. k vec is calculated separately for patents in different fields. k The similarity to these patent vectors is calculated using the following formula: The similarity scores are then sorted from highest to lowest to obtain a "Design Requirements - Patents" similarity list.

[0096] Step 13: Push the top-ranked patents in the "Design Requirements - Patents" similarity list to the user. In this embodiment, the three patents with the highest functional or technical similarity are pushed, for a total of six patents. Their titles and fields are shown in Table 3.

[0097]

[0098]

[0099] The key point of this invention is:

[0100] A patent feature space based on <domain (D), function (F), and technical term (T)> was constructed, and these three features were used to train the content embedding vector of the patent.

[0101] A patent embedding vector learning model integrating Doc2vec and GAT models was constructed. The patent abstract content and patent cross-citation relationship were used as training content. The training model learned embedding vectors that can represent patent semantics in high-dimensional space.

[0102] During the patent recommendation stage, the Doc2vec model can be used to transform the user's input knowledge requirements into a vector, which can then be used to calculate with the patent embedding vector.

[0103] Any process or method described in the flowcharts of this invention or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, achievable on any computer-readable medium for use by an instruction execution system, apparatus, or device. The computer-readable medium can be any medium containing a program for storage, communication, propagation, or transmission for use by an execution system, apparatus, or device, including read-only memory, magnetic disks, or optical disks.

[0104] In the description of this specification, references to terms such as "embodiment," "example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, those skilled in the art can combine or combine the different embodiments or examples described in this specification and the features therein without causing contradiction.

[0105] While embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions, and alterations to the above embodiments within the scope of the present invention.

Claims

1. A cross-domain patent recommendation method integrating Doc2vec and GAT, characterized in that, Includes the following steps: Step S100: Data collection; Collect the publication number, title, abstract and citation relationship of the patent from the open source patent platform, and preprocess the text content; Step S200: Model training; Use Doc2vec to learn the patent content and train the GAT model to fuse the patent content and citation relationships; Step S300: Search and recommend; input the design requirements into the trained model, vectorize the design requirements, and obtain the recommended cross-domain patents; Step S200 includes: Step S2: Use the BERT model to perform named entity recognition on the patent abstract; Step S3: Determine whether the collected patent abstracts indicate the field, function, or technical information; Step S4: If the patent does not contain any information in the field, function, or technology, it means that the patent is not related to the innovative design of complex products, and the patent will not be trained for the time being. Step S5: If the patent is marked with field, functional information, or technical terms, then use Doc2vec to learn the patent content; Step S6: Construct a patent cross-citation network based on patent citation relationships; Step S7: Embed the patent citation network information into vectors along with the patent's field, function, and technical terms. Three GAT models are input for parameter tuning and training to update the patent embedding vectors; Step S8: For each patent By splicing, we get < A vector is an embedded vector that contains domain information, functional information, technical terminology information, and reference information.

2. The patent cross-domain recommendation method integrating Doc2vec and GAT according to claim 1, characterized in that, In step S2, firstly, high-quality named entity recognition data in a certain field is manually annotated. The field, functional information, and technical terms in the patent abstract are labeled as D, F, and T, respectively; where D represents the field, F represents the function, and T represents the technical term. The manually annotated named entity recognition dataset is then input into the BERT model for parameter tuning and learning. The main parameters to be adjusted are sequence length, batch size, and number of iterations. After that, the trained BERT model is used to perform named entity recognition on the unannotated patent abstracts and annotate the field, functional information, or technical term information.

3. The patent cross-domain recommendation method integrating Doc2vec and GAT according to claim 1, characterized in that, In step S5, the specific method for learning the patent content using Doc2vec is as follows: merge the text content of the patent title and abstract to train three Doc2vec models, and train only one Doc2vec model using the patent title, for a total of four Doc2vec models. At the same time, three content embedding vectors are also obtained for each patent. This forms a set of content embedding vectors. and a title embedding vector The corresponding title embedding vector set The parameters that need to be adjusted during the training of the Doc2vec model are the sliding window size, learning rate, and number of iterations.

4. The patent cross-domain recommendation method integrating Doc2vec and GAT according to claim 1, characterized in that, In step S6, a patent citation network is constructed based on patent citation relationships; if there is a citation between two patents, the two are connected by an edge in the network to form a patent citation network of <initial patent, cite, first-level citation patent, cite, second-level citation patent>.

5. The patent cross-domain recommendation method integrating Doc2vec and GAT according to claim 1, characterized in that, In step S7, the patent embedding vector is updated so that the vector of each patent is updated. They respectively integrate the reference relationship information, and are represented as .

6. The patent cross-domain recommendation method integrating Doc2vec and GAT according to claim 1, characterized in that, Step S300 includes: Step S9: The user inputs the design requirements to be addressed; Step S10: Vectorize the design requirements; use the Doc2vec model to transform the user-input design requirements into a design requirement vector, represented as follows: , r represents demand; Step S11: Identify the patents that represent the design requirements; Step S12: Retrieve patents that are technically or functionally related but in different fields from the trained patent embedding vector model; Step S13: Push the patents ranked first in the "Design Requirements - Patents" similarity list to the user.

7. The patent cross-domain recommendation method integrating Doc2vec and GAT according to claim 6, characterized in that, In step S11, respectively at Similarity calculations are performed, and the top few patents with the highest similarity are pushed to the user. After reading them, the user selects one patent as a replacement for their design requirements. The vector representation of this patent is as follows: .

8. The patent cross-domain recommendation method integrating Doc2vec and GAT according to claim 6, characterized in that, In step S12, the specific operation method is as follows: retrieve the results related to... Patents in different fields are calculated separately. The similarity between these patent vectors is calculated and sorted from highest to lowest similarity to obtain a "Design Requirement - Patent" similarity list.

9. A cross-domain patent recommendation system integrating Doc2vec and GAT, characterized in that, The system is used to implement the method according to any one of claims 1-8, and the system comprises: The data acquisition module is used to acquire patent information, which includes the patent publication number, title, abstract, and citation relationships. Model training module: Uses Doc2vec to learn from patent content and trains the GAT model to integrate patent content and citation relationships; Search and recommendation module: Inputs design requirements into the trained model, vectorizes the design requirements, and obtains recommended cross-domain patents.

Images