Product-based intellectual property search and classification method, device, equipment and medium

Abstract

The present invention relates to the field of network information processing technology, and discloses a product-based intellectual property retrieval classification method, device, equipment and medium, the method comprising: combing trademark and patent application data and corresponding NICE, LOC, CPC classifications, formulating classification standards after evaluating cross-correlation, building a database, and preprocessing the selected trademark application business scope description and patent text in the database based on the NLP process; performing multi-stage feature extraction after preprocessing, combining classification standards, and using document-level features to perform training and fitting on the classification model; classifying and predicting the obtained product title or product description to be tested, recalling the relevant trademark and patent list as candidate infringement items, and displaying the business scope, NICE and corresponding patent LOC and CPC classification of the trademark matched by the product. The present invention can reduce the difficulty of merchants extracting product keywords and product classifications, and can quickly and accurately match products with the business scope and patent classification applicable to trademarks.

Description

Technical Field

[0001] The present invention relates to the technical field of network information processing, and in particular to a product-based intellectual property retrieval classification method, device, equipment and medium. Background Art

[0002] In the e-commerce industry, especially for foreign e-commerce platforms, it is necessary to classify the products they sell. For most e-commerce store owners, due to the lack of professional intellectual property search and keyword extraction capabilities, they are not clear about the business scope of the trademark applicable to their products, and are not clear about the LOC and CPC classifications corresponding to the patents that may be applied for. In particular, it is difficult for non-professionals to search for patents and trademarks related to the products, and they are prone to the risk of infringement lawsuits in the later stage. Summary of the invention

[0003] In view of the above technical problems, the present invention provides a product-based intellectual property retrieval classification method, device, equipment and medium to solve the problems mentioned in the above background technology.

[0004] Other features and advantages of the present disclosure will become apparent from the following detailed description, or may be learned in part by the practice of the present disclosure.

[0005] According to one aspect of the present invention, a product-based intellectual property retrieval classification method is proposed, the method comprising:

[0006] Sort out trademark and patent application data and the corresponding NICE, LOC, CPC classifications, evaluate the intersection of NICE, LOC, CPC classifications, and then formulate NICE, LOC, and CPC learning classification standards, build a database, and perform data preprocessing on the business scope descriptions and patent texts of trademark applications selected from the database based on the NLP process. The preprocessing includes segmenting the patent text and the business scope description into paragraphs and obtaining the feature matrix of the paragraphs;

[0007] Perform multi-stage feature extraction on the preprocessed patent text and the business scope description. During the extraction, the features of the paragraphs are first summarized, and then the features of the paragraphs are aggregated to generate document-level features. In combination with the NICE, LOC, and CPC learning classification standards, the document-level features are used to perform training and fitting on the classification model;

[0008] The obtained title or description of the product to be tested is input into the classification model for classification prediction, and a list of related trademarks and patents is recalled as candidate infringement items based on the classification results, and the business scope, NICE and corresponding patent LOC and CPC classification of the trademark matching the product are displayed.

[0009] Furthermore, after evaluating the intersection of NICE, LOC, and CPC classifications, NICE, LOC, and CPC learning classification standards are formulated, including:

[0010] Redefine the cross-classification;

[0011] Define common categories for e-commerce products and delete categories for products that are not within the defined scope.

[0012] Furthermore, in the preprocessing, the patent text and the business scope description are segmented into paragraphs with a fixed length, the title and abstract of the patent text are used to segment the paragraphs, a complete sentence does not span two paragraphs, and paragraphs with insufficient length are supplemented with filling markers, and each word in the paragraph is converted into a D-dimensional feature vector to obtain the feature matrix.

[0013] Furthermore, the preprocessed patent text and the business scope description are subjected to multi-stage feature extraction, including:

[0014] Encoding the feature matrix into paragraph-level features of a fixed size based on one of a convolutional neural network, a recurrent neural network, or an attention mechanism, so that the context of the feature matrix is ​​summarized only based on words and sentences within the paragraph;

[0015] A bidirectional GRU encoder is used to generate a weighted representation of each paragraph-level feature through a linear combination of the encoder state and the decoder device, and the weighted sum is performed to generate the final document-level feature, or a stacked transformer encoder is used to mark all the paragraph-level features through classification tags, and the tags are used at the end of each layer of the stacked transformer encoder to summarize the paragraph-level features of the corresponding layer to generate a comprehensive document-level feature.

[0016] Furthermore, the feature matrix is ​​encoded into paragraph-level features of a fixed size based on an attention mechanism, including:

[0017] Based on stacked transformer encoders, the paragraph-level features are extracted in a hierarchical manner.

[0018] Furthermore, the combining of NICE, LOC, and CPC learning classification standards and using the document-level features to perform training and fitting on the classification model include:

[0019] Generate the corresponding prediction weight vectors according to the label information of the parent category and subcategory of NICE, LOC, and CPC;

[0020] When the number of parent categories and subcategories is different, the vector dimension of the parent category's label information is expanded to make it have the same size as the vector dimension of the subcategory's label information, and the parent category's vector weight is broadcasted to the expanded vector to obtain a comprehensive prediction vector combining the parent category and the subcategory;

[0021] The comprehensive prediction vector is used as a label and combined with the document-level features to train the classification model.

[0022] According to a second aspect of the present disclosure, there is provided a product-based intellectual property retrieval and classification device, comprising:

[0023] A preprocessing module is used to sort out trademark and patent application data and the corresponding NICE, LOC, and CPC classifications, formulate NICE, LOC, and CPC learning classification standards after evaluating the intersection of NICE, LOC, and CPC classifications, build a database, and perform data preprocessing on the business scope description and patent text of the trademark application selected in the database based on the NLP process. The preprocessing includes segmenting the patent text and the business scope description into paragraphs and obtaining a feature matrix of the paragraphs;

[0024] A model training module is used to perform multi-stage feature extraction on the pre-processed patent text and the business scope description. During the extraction, the features of the paragraphs are first summarized, and then the features of the paragraphs are aggregated to generate document-level features. In combination with the NICE, LOC, and CPC learning classification standards, the document-level features are used to perform training and fitting on the classification model;

[0025] The retrieval module is used to input the obtained product title or product description to the classification model for classification prediction, recall the relevant trademark and patent list as candidate infringement items according to the classification results, and display the business scope, NICE and corresponding patent LOC and CPC classification of the trademark matching the product.

[0026] According to a third aspect of the present disclosure, a product-based intellectual property retrieval and classification device is provided, the device comprising:

[0027] A processor; and a memory arranged to store computer executable instructions, which, when executed, cause the processor to perform the above-mentioned product-based intellectual property search classification method.

[0028] According to a fourth aspect of the present disclosure, there is provided a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the product-based intellectual property retrieval classification method as described above.

[0029] The technical solution disclosed in this disclosure has the following beneficial effects:

[0030] It reduces the difficulty for merchants to extract product keywords and product classifications, and helps non-professionals to quickly and accurately match products to the business scope of the trademark applicable to such products, as well as the LOC and CPC classifications corresponding to possible patent applications; the model used to classify products combines paragraph-level features and document-level features during training, making it possible to classify and retrieve long document patents. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 A flowchart of a product-based intellectual property search and classification method in an embodiment of this specification;

[0032] Figure 2 is a calculation flow chart of a bidirectional GRU encoder in an embodiment of this specification;

[0033] Figure 3 This is a calculation flow chart of the stacked transformer encoder in the embodiment of this specification;

[0034] Figure 4 It is a structural block diagram of a product-based intellectual property search and classification device in an embodiment of this specification;

[0035] Figure 5 It is a terminal device for implementing the product-based intellectual property search and classification method in the embodiments of this specification;

[0036] Figure 6 It is a computer-readable storage medium for storing a product-based intellectual property search classification method in an embodiment of this specification. DETAILED DESCRIPTION

[0037] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in a variety of forms and should not be construed as being limited to the examples set forth herein; on the contrary, these embodiments are provided so that the present disclosure will be more comprehensive and complete, and the concepts of the example embodiments are fully conveyed to those skilled in the art. The described features, structures, or characteristics may be combined in one or more embodiments in any suitable manner. In the following description, many specific details are provided to provide a full understanding of the embodiments of the present disclosure. However, those skilled in the art will appreciate that the technical solutions of the present disclosure may be practiced while omitting one or more of the specific details, or other methods, components, devices, steps, etc. may be adopted. In other cases, known technical solutions are not shown or described in detail to avoid obscuring various aspects of the present disclosure.

[0038] In addition, the accompanying drawings are only schematic illustrations of the present disclosure. The same reference numerals in the figures represent the same or similar parts, and thus their repeated description will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities, which do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software form, or implemented in one or more hardware modules or integrated circuits, or implemented in different networks and / or processor devices and / or microcontroller devices.

[0039] The present invention provides a product classification method based on product intellectual property rights. Figure 1 As shown, it is a flow chart of a product-based intellectual property search and classification method provided by an embodiment of the present invention. The method can be applied to electronic devices such as personal computers and servers. The method can be executed by a device, and the device can be implemented by software and / or hardware. The method can specifically include the following steps S101 to S103:

[0040] In step S101, the trademark and patent application data and the corresponding NICE, LOC, CPC classifications are sorted out, the intersection of NICE, LOC, CPC classifications is evaluated, and then the NICE, LOC and CPC learning classification standards are formulated, and a database is constructed. Based on the NLP process, data preprocessing is performed on the business scope description and patent text of the trademark application selected in the database. The preprocessing includes segmenting the patent text and the business scope description into paragraphs and obtaining a feature matrix of the paragraphs.

[0041] Among them, NICE is the Nice Classification. The Nice Classification data is derived from the registered trademark information of domestic and foreign trademark offices. The registered trademark information includes trademark name, classification Nice Classification, registered specific product list, registration time, right holder, related business description, etc. During sorting, the Nice Classification, registered specific product list and specific business description are cleaned and extracted. CPC classification data is derived from the registered invention and utility model patent information of domestic and foreign patent offices. These patent information generally include patent number, publication number, publication date, applicant, patent name, abstract, description, claims, instructions, drawings, CPC classification, etc. During sorting, the following are cleaned and extracted: patent name, abstract, claims, instructions and CPC classification number as data, among which patent name, abstract, claims and instructions constitute patent text; LOC classification data is similarly derived from the registered appearance patent information of the patent office, and the patent name, LOC classification number and pictures need to be extracted.

[0042] After sorting out the information required for classification, it is necessary to re-formulate the NICE, LOC and CPC learning classification standards, redefine the cross-classification, define the common classifications for e-commerce products, and delete the classifications of products that are not within the definition range. For example, in the Nice classification, doll clothes and human clothes belong to two Nice classifications, which will reduce the performance for model training, so the classification standards need to be redefined; for example, in the three classification systems, there are many products that e-commerce does not use, such as the categories of nuclear industry, pharmaceutical manufacturing, shampoo formula, books and periodicals, etc. These can be removed to simplify the subsequent classification retrieval work. The organized trademark and patent data and their classification information are stored in the database, which will serve as the data source for subsequent retrieval.

[0043] After the database is established, representative data of patents and trademarks of various categories can be selected for preprocessing. In the preprocessing, the patent text and the business scope description are divided into paragraphs with a fixed length. The title and abstract of the patent text are used to divide the paragraphs. A complete sentence does not span two paragraphs. Paragraphs with insufficient length are supplemented with filling marks. Each word in the paragraph is converted into a D-dimensional feature vector to obtain the feature matrix.

[0044] As an explanation, in model training, the more input words are used, the better the performance is. However, various existing models are difficult to effectively handle long inputs. Long inputs will greatly increase computing costs and memory consumption, causing a significant burden on business operations. In order to balance this problem, the present invention uses paragraphs as the basic unit of input data. 1 ,T 2 ,...,T N ] and the classification label L = [L 1 ,L 2 ,...,L N ], before inputting data, each patent text and business scope description is segmented into paragraphs with a fixed word length I, and all words or words are embedded into D-dimensional features. In order to prevent the continuity of the sentence from being torn, a sentence can be designed not to be divided into two consecutive paragraphs, and at the end of the paragraph, as many single filler words as possible are used to fill in the length. In order to correct that all important features of the patent can be used for analysis, the title and abstract of the patent also need to be regarded as paragraphs. The feature matrix is ​​represented as P i ∈R D×l .

[0045] In step S102, multi-stage feature extraction is performed on the preprocessed patent text and the business scope description. During the extraction, the features of the paragraphs are first summarized, and then the features of the paragraphs are aggregated to generate document-level features. The classification model is trained and fitted using the document-level features in combination with the NICE, LOC, and CPC learning classification standards.

[0046] Among them, in the feature extraction stage of the paragraph, the content of each paragraph can be analyzed to extract features related to the paragraph, such as using word embedding to represent words in the paragraph, or identifying key phrases and technical terms in the paragraph. After extracting the features at the paragraph level, these features are summarized to capture the main information of the paragraph. Techniques such as average pooling or maximum pooling can be used to obtain representative feature vectors of the paragraph. When aggregating these paragraph-level features to generate document-level features, the feature vectors of all paragraphs can be connected to form a long vector, such as using an attention mechanism to weight the features of each paragraph to highlight the paragraphs that are more important for the overall understanding of the document, and applying encoders such as Transformer to process paragraph features to generate comprehensive document-level feature representations. After generating document-level features, the training of the classification model is guided by learning classification standards such as NICE, LOC, and CPC.

[0047] In step S103, the obtained title or product description of the product to be tested is input into the classification model for classification prediction, and a list of related trademarks and patents is recalled as candidate infringement items based on the classification results, and the business scope, NICE and corresponding patent LOC and CPC classification of the trademark matching the product are displayed.

[0048] Among them, the obtained product title or product description can be input into the classification as a paragraph, and the matching degree between the paragraph and document-level features is used to make classification predictions. The patent and trademark information retrieved from the database is displayed, and the most likely NICE, LOC and CPC classifications are displayed at the same time, so that users can clearly see the intellectual property rights related to their own products on the store page of the e-commerce platform.

[0049] In one embodiment, the multi-stage feature extraction of the pre-processed patent text and the business scope description specifically includes:

[0050] The feature matrix is ​​encoded into a paragraph-level feature of a fixed size based on one of a convolutional neural network, a recurrent neural network or an attention mechanism, so that the context of the feature matrix is ​​summarized only based on the words and sentences in the paragraph. A bidirectional GRU encoder is used to generate a weighted representation of each paragraph-level feature through a linear combination of the encoder state and the decoder device, and the weighted sum is used to generate the final document-level feature, or a stacked transformer encoder is used to mark all the paragraph-level features through classification tags, and the tags are used at the end of each layer of the stacked transformer encoder to summarize the paragraph-level features of the corresponding layer to generate a comprehensive document-level feature.

[0051] In this embodiment, each fixed-size paragraph feature matrix P i ∈R D×l The encoding is paragraph-level features This can be used to summarize feature matrices based on words and sentences, using only the context of the paragraph itself without considering other adjacent paragraphs. Similarly, all words or terms in the paragraph will be embedded into a predefined feature of length I. This can be achieved using convolutional neural networks, recurrent neural networks, or attention mechanisms.

[0052] When using convolutional neural network, the calculation formula F(P i )=W θ P i +b, i is the index of the paragraph position, W θ , b represent the convolution weight parameter and bias respectively. i is a fixed-size feature matrix and can therefore be computed directly.

[0053] When using a recurrent neural network, the calculation W θ and W φ , b represent the weight parameters and bias of the learnable recurrent neural network respectively.

[0054] When using the attention mechanism, the paragraph-level features can be extracted in a hierarchical manner based on a stacked transformer encoder. Specifically, Among them, Q, K, V represent the query, key and value of the transformer encoder respectively, and the features are obtained by concatenating the [CLS] tag and P i To embed, as Q i =K i =V i ={[CLS],P i}, where {,} represents concatenation, L is the total number of transformer encoders, l is the index of the encoder, and L and s determine how many encoder layers are used to summarize the paragraph. This contextual word embedding method is different from the above-mentioned convolutional neural network and recurrent neural network.

[0055] Since the patent text has many paragraphs, after the paragraph-level features are encoded and calculated, these paragraph-level features need to be summarized together, which can be done through a bidirectional GRU encoder or a stacked transformer encoder, that is, an additive attention mechanism or a productive attention mechanism.

[0056] When using a bidirectional GRU encoder, the following formula is calculated:

[0057]

[0058] Among them, α i For each paragraph-level feature The weight of the context vector u i and the state vector u s The context vector u is calculated by the dot product of i Represents the summary paragraph-level features, which can be obtained from the hidden vector Export in, hide vectors It is composed of bidirectional GRU encoders connected, such as Figure 2 As shown in Figure 2, a bidirectional GRU encoder is used to mine paragraph sequences, and the paragraph features are summarized through a linear combination of the encoder state and the decoder state. This is not easily affected by the gradient vanishing problem and is more suitable for long patent texts.

[0059] When using a stacked transformer encoder, the following formula is calculated:

[0060]

[0061] Where l is the encoder A D The number of, Q, K, V are the query, key and value of the transformer encoder respectively. To summarize the paragraph-level features, a symmetric operation of injecting the maximum, minimum and average values ​​can be used, as well as using classification tags [DCLS]. In the case of tags, the features are embedded in the concatenated tags [DCLS], then can be calculated as By adding a token before the input paragraph-level features, the paragraph-level features can be summarized on each encoder stack layer to obtain document-level features, such as Figure 3 As shown in Figure 2. Based on the stacked transformer encoder, all input paragraph-level features can be processed simultaneously through dot product operations, which improves processing efficiency.

[0062] In one embodiment, the classification standards are learned in combination with NICE, LOC, and CPC, and the document-level features are used to perform training and fitting on the classification model, including: generating corresponding prediction weight vectors according to the label information of the parent category and the child category of NICE, LOC, and CPC; when the number of parent categories and child categories is different, expanding the vector dimension of the label information of the parent category to make it have the same size as the vector dimension of the label information of the child category, and broadcasting the vector weight of the parent category to the expanded vector to obtain a comprehensive prediction vector combining the parent category and the child category; using the comprehensive prediction vector as a label, combined with the document-level features, to train the classification model.

[0063] Among them, the developed NICE, LOC, and CPC classification standards can be used to enhance the prediction reliability by comparing the interdependencies between parent categories and subcategories, which can be obtained by calculating the following formula:

[0064] v′=J()=BroadCast(v p )×v c ;

[0065] in, and are the prediction weight vectors of the parent category and the child category respectively. Since the number of categories in NICE, LOC, and CPC is different, the vector dimension of the parent category needs to be expanded to the size of the child category, and then the vector weight of the parent category is broadcast to the expanded favorite. The weights of the parent category and the child category are represented by v p =W p v+b p and v c =W c v+b c It is calculated that W and b are the weight matrix and bias that need to be learned respectively.

[0066] Finally, through the fully connected network, the distribution of multiple parent categories and subcategories is generated. In addition, the optimization for multi-label and multi-class classification can be calculated as follows:

[0067] l(x,y)=L={l 1 ,...,l N},l n =-[y n ·logσ(x n )+(1-y n )·log(1-σ(x n ))];

[0068] Here, the training loss is coordinated based on the target y, the predicted label x, and the sigmoid function σ.

[0069] Based on the same idea, Figure 4 As shown, a product-based intellectual property search and classification device is provided, comprising:

[0070] A preprocessing module 401 is used to sort out trademark and patent application data and corresponding NICE, LOC, CPC classifications, formulate NICE, LOC and CPC learning classification standards after evaluating the intersection of NICE, LOC and CPC classifications, build a database, and perform data preprocessing on the business scope description and patent text of the trademark application selected in the database based on the NLP process, wherein the preprocessing includes segmenting the patent text and the business scope description into paragraphs and obtaining a feature matrix of the paragraphs;

[0071] Model training module 402, for performing multi-stage feature extraction on the pre-processed patent text and the business scope description, first summarizing the features of the paragraphs during extraction, and then aggregating the features of the paragraphs to generate document-level features, combining NICE, LOC, CPC learning classification standards, and using the document-level features to perform training and fitting on the classification model;

[0072] The retrieval module 403 is used to input the obtained product title or product description to the classification model for classification prediction, recall the relevant trademark and patent list as candidate infringement items according to the classification results, and display the business scope, NICE and corresponding patent LOC and CPC classification of the trademark matching the product.

[0073] The above-mentioned product-based intellectual property retrieval and classification device reduces the difficulty for merchants to extract product keywords and product classifications, and helps non-professionals to quickly and accurately match products with the business scope of the trademark applicable to such products, as well as the LOC and CPC classifications corresponding to the patents that may be applied for; the model used to classify products combines paragraph-level features and document-level features during training, so that long document patents can be classified and retrieved.

[0074] The specific details of each module / unit in the above-mentioned device have been described in detail in the implementation method of the method part. The undisclosed details can be found in the implementation method of the method part, so they will not be repeated here.

[0075] Based on the same idea, the embodiment of this specification also provides a product-based intellectual property retrieval and classification device, such as Figure 5 shown.

[0076] The product-based intellectual property retrieval and classification device may be the terminal device or server provided in the above embodiment.

[0077] The product-based intellectual property search and classification device may have relatively large differences due to different configurations or performances, and may include one or more processors 501, memory 502 and a bus, and the memory 502 may store one or more storage applications or data. Among them, the memory 502 may include a readable medium in the form of a volatile storage unit, such as a random access storage unit (RAM) and / or a cache storage unit, that is, for example, a plug-in mobile hard disk, a smart memory card (SmartMediaCard, SMC), a secure digital (SecureDigital, SD) card, a flash card (FlashCard), etc. equipped on an electronic device, and may further include a read-only storage unit. The application stored in the memory 502 may include one or more program modules (not shown in the figure), such program modules include but are not limited to: an operating device, one or more application programs, other program modules and program data, each of these examples or a certain combination may include the implementation of a network environment. Furthermore, the processor 501 may be configured to communicate with the memory 502 to execute a series of computer executable instructions in the memory 502 on the product-based intellectual property search and classification device. The product-based intellectual property search and classification device may also include one or more power supplies 503, one or more wired or wireless network interfaces 504, one or more I / O interfaces (input and output interfaces) 505, one or more external devices 506 (such as a keyboard), and may also communicate with one or more devices that enable a user to interact with the device, and / or communicate with any device that enables the device to communicate with one or more other computing devices (such as a router, a network switch, etc.). Such communication may be performed through the I / O interface 505. In addition, the device may also communicate with one or more networks (such as a local area network (LAN)) through the wired or wireless interface 504.

[0078] Figure 6 Only a product-based intellectual property search and classification device with components is shown, and those skilled in the art can understand that Figure 6 The structure shown does not constitute a limitation on the product-based intellectual property search and classification device, and may include fewer or more components than shown in the figure, or combine certain components, or arrange the components differently.

[0079] Specifically in this embodiment, the product-based intellectual property search and classification device includes a memory and one or more programs, wherein the one or more programs are stored in the memory, and the one or more programs may include one or more modules, and each module may include a series of computer executable instructions in the product-based intellectual property search and classification device, and the one or more programs are configured to be executed by one or more processors, including the following computer executable instructions:

[0080] Sort out trademark and patent application data and the corresponding NICE, LOC, CPC classifications, evaluate the intersection of NICE, LOC, CPC classifications, and then formulate NICE, LOC, and CPC learning classification standards, build a database, and perform data preprocessing on the business scope descriptions and patent texts of trademark applications selected from the database based on the NLP process. The preprocessing includes segmenting the patent text and the business scope description into paragraphs and obtaining the feature matrix of the paragraphs;

[0081] Perform multi-stage feature extraction on the preprocessed patent text and the business scope description. During the extraction, the features of the paragraphs are first summarized, and then the features of the paragraphs are aggregated to generate document-level features. In combination with the NICE, LOC, and CPC learning classification standards, the document-level features are used to perform training and fitting on the classification model;

[0082] The obtained title or description of the product to be tested is input into the classification model for classification prediction, and a list of related trademarks and patents is recalled as candidate infringement items based on the classification results, and the business scope, NICE and corresponding patent LOC and CPC classification of the trademark matching the product are displayed.

[0083] Based on the same idea, the exemplary embodiment of the present invention also provides a computer-readable storage medium, on which a program product capable of implementing the above method of this specification is stored. In some possible implementations, various aspects of the present disclosure can also be implemented in the form of a program product, which includes a program code. When the program product is run on a terminal device, the program code is used to enable the terminal device to execute the steps according to various exemplary embodiments of the present disclosure described in the above "Exemplary Method" section of this specification.

[0084] refer to Figure 6As shown, a program 600 for implementing the above method according to an exemplary embodiment of the present disclosure is described, which can adopt a portable compact disk read-only memory (CD-ROM) and include program code, and can be run on a terminal device, such as a personal computer. However, the program product of the present disclosure is not limited thereto, and in this document, a readable storage medium can be any tangible medium containing or storing a program, which can be used by or in combination with an instruction execution device, apparatus, or device.

[0085] The program product may use any combination of one or more readable media. The readable medium may be a readable signal medium or a readable storage medium. The readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor device, device or component, or any combination of the above. More specific examples of readable storage media (a non-exhaustive list) include: an electrical connection with one or more wires, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above.

[0086] Computer readable signal media may include data signals propagated in baseband or as part of a carrier wave, wherein readable program code is carried. Such propagated data signals may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. Readable signal media may also be any readable medium other than a readable storage medium, which may send, propagate, or transmit a program for use by or in conjunction with an instruction execution device, apparatus, or device.

[0087] Program code for performing the operations of the present disclosure may be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java, C++, CSS, HTML, etc., and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on the user computing device, partially on the user device, as a separate software package, partially on the user computing device and partially on a remote computing device, or entirely on a remote computing device or server. In cases involving a remote computing device, the remote computing device may be connected to the user computing device through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0088] Through the description of the above implementation, it is easy for those skilled in the art to understand that the example implementation described here can be implemented by software, or by software combined with necessary hardware. Therefore, the technical solution according to the implementation of the present disclosure can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a USB flash drive, a mobile hard disk, etc.) or on a network, including several instructions to enable a computing device (which can be a personal computer, a server, a terminal device, or a network device, etc.) to execute the method according to the exemplary implementation of the present disclosure.

[0089] In addition, the above-mentioned figures are only schematic illustrations of the processes included in the method according to the exemplary embodiment of the present disclosure, and are not intended to be limiting. It is easy to understand that the processes shown in the above-mentioned figures do not indicate or limit the time sequence of these processes. In addition, it is also easy to understand that these processes can be performed synchronously or asynchronously, for example, in multiple modules.

[0090] It should be noted that, although several modules or units of the device for action execution are mentioned in the above detailed description, this division is not mandatory. In fact, according to the exemplary embodiments of the present disclosure, the features and functions of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided into multiple modules or units to be embodied.

[0091] Those skilled in the art will readily appreciate other embodiments of the present disclosure after considering the specification and practicing the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the present disclosure that follow the general principles of the present disclosure and include common knowledge or customary technical means in the art that are not disclosed in the present disclosure. The specification and embodiments are to be considered exemplary only, and the true scope and spirit of the present disclosure are indicated by the claims.

[0092] It should be understood that the present disclosure is not limited to the exact structures that have been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present disclosure is limited only by the appended claims.

Claims

1. A product-based intellectual property search and classification method, characterized in that: The method comprises: Sort out trademark and patent application data and the corresponding NICE, LOC, CPC classifications, evaluate the intersection of NICE, LOC, CPC classifications, and then formulate NICE, LOC and CPC learning classification standards, build a database, and perform data preprocessing on the selected trademark application business scope description and patent text in the database based on the NLP process. The preprocessing includes segmenting the patent text and the business scope description into paragraphs with a fixed length. The title and abstract of the patent text are used to segment paragraphs. A complete sentence does not span two paragraphs. Paragraphs with insufficient length are supplemented with filling tags. Each word in the paragraph is converted into a D-dimensional feature vector to obtain a feature matrix. A multi-stage feature extraction is performed on the preprocessed patent text and the business scope description. During the extraction, the feature matrix is ​​encoded into a paragraph-level feature of a fixed size based on a convolutional neural network, a recurrent neural network or an attention mechanism, so that the context of the feature matrix is ​​summarized only based on the words and sentences in the paragraph; a bidirectional GRU encoder is used to generate a weighted representation of each paragraph-level feature through a linear combination of an encoder state and a decoder device, and a weighted sum is performed to generate a final document-level feature, or a stacked transformer encoder is used to mark all the paragraph-level features through classification marking, and then a weighted sum is generated in each paragraph-level feature. The end of a stacked transformer encoder uses the tag to summarize the paragraph-level features of the corresponding layer to generate a comprehensive document-level feature; according to the label information of the parent category and the subcategory of NICE, LOC, and CPC, the corresponding prediction weight vectors are generated respectively; when the number of parent categories and subcategories is different, the vector dimension of the label information of the parent category is expanded to make it have the same size as the vector dimension of the label information of the subcategory, and the vector weight of the parent category is broadcasted to the expanded vector to obtain a comprehensive prediction vector combining the parent category and the subcategory, and the comprehensive prediction vector is used as a label, combined with the document-level feature, to train a classification model; The obtained title or description of the product to be tested is input into the classification model for classification prediction, and a list of related trademarks and patents is recalled as candidate infringement items based on the classification results, and the business scope, NICE and corresponding patent LOC and CPC classification of the trademark matching the product are displayed.

2. The product-based intellectual property search and classification method according to claim 1 is characterized in that: The above assessment of the intersection of NICE, LOC, and CPC classifications led to the development of NICE, LOC, and CPC learning classification standards, including: Redefine the cross-classification; Define common categories for e-commerce products and delete categories for products that are not within the defined scope.

3. The product-based intellectual property search and classification method according to claim 1 is characterized in that: The feature matrix is ​​encoded into a fixed-size paragraph-level feature based on the attention mechanism, including: Based on stacked transformer encoders, the paragraph-level features are extracted in a hierarchical manner.

4. A product-based intellectual property search and classification device, characterized in that: include: A preprocessing module is used to sort out trademark and patent application data and the corresponding NICE, LOC, CPC classifications, formulate NICE, LOC and CPC learning classification standards after evaluating the intersection of NICE, LOC and CPC classifications, build a database, and perform data preprocessing on the business scope description and patent text of the trademark application selected in the database based on the NLP process. The preprocessing includes segmenting the patent text and the business scope description into paragraphs with a fixed length. The title and abstract of the patent text are used to segment paragraphs. A complete sentence does not span two paragraphs. Paragraphs with insufficient length are supplemented with filling tags. Each word in the paragraph is converted into a D-dimensional feature vector to obtain a feature matrix; A model training module, for performing multi-stage feature extraction on the preprocessed patent text and the business scope description, during which the feature matrix is ​​encoded into paragraph-level features of a fixed size based on one of a convolutional neural network, a recurrent neural network, or an attention mechanism, so that the context of the feature matrix is ​​summarized based only on words and sentences within a paragraph; A bidirectional GRU encoder is used to generate a weighted representation of each paragraph-level feature through a linear combination of the encoder state and the decoder device, and a weighted sum is performed to generate a final document-level feature, or a stacked transformer encoder is used to mark all the paragraph-level features through classification tags, and the tags are used at the end of each layer of the stacked transformer encoder to summarize the paragraph-level features of the corresponding layer to generate a comprehensive document-level feature; according to the label information of the parent category and the subcategory of NICE, LOC, and CPC, the corresponding prediction weight vectors are generated respectively, and when the number of parent categories and subcategories is different, the vector dimension of the label information of the parent category is expanded to make it have the same size as the vector dimension of the label information of the subcategory, and the vector weight of the parent category is broadcasted to the expanded vector to obtain a comprehensive prediction vector combining the parent category and the subcategory, and the comprehensive prediction vector is used as a label, combined with the document-level features, to train a classification model; The retrieval module is used to input the obtained product title or product description to the classification model for classification prediction, recall the relevant trademark and patent list as candidate infringement items according to the classification results, and display the business scope, NICE and corresponding patent LOC and CPC classification of the trademark matching the product.

5. A product-based intellectual property retrieval and classification device, characterized in that: The device comprises: a processor; and a memory arranged to store computer executable instructions which, when executed, cause the processor to: Sort out trademark and patent application data and the corresponding NICE, LOC, CPC classifications, evaluate the intersection of NICE, LOC, CPC classifications, and then formulate NICE, LOC and CPC learning classification standards, build a database, and perform data preprocessing on the selected trademark application business scope description and patent text in the database based on the NLP process. The preprocessing includes segmenting the patent text and the business scope description into paragraphs with a fixed length. The title and abstract of the patent text are used to segment paragraphs. A complete sentence does not span two paragraphs. Paragraphs with insufficient length are supplemented with filling tags. Each word in the paragraph is converted into a D-dimensional feature vector to obtain a feature matrix. A multi-stage feature extraction is performed on the preprocessed patent text and the business scope description. During the extraction, the feature matrix is ​​encoded into a paragraph-level feature of a fixed size based on a convolutional neural network, a recurrent neural network or an attention mechanism, so that the context of the feature matrix is ​​summarized only based on the words and sentences in the paragraph; a bidirectional GRU encoder is used to generate a weighted representation of each paragraph-level feature through a linear combination of an encoder state and a decoder device, and a weighted sum is performed to generate a final document-level feature, or a stacked transformer encoder is used to mark all the paragraph-level features through classification marking, and then a weighted sum is generated in each paragraph-level feature. The end of a stacked transformer encoder uses the tag to summarize the paragraph-level features of the corresponding layer to generate a comprehensive document-level feature; according to the label information of the parent category and the subcategory of NICE, LOC, and CPC, the corresponding prediction weight vectors are generated respectively; when the number of parent categories and subcategories is different, the vector dimension of the label information of the parent category is expanded to make it have the same size as the vector dimension of the label information of the subcategory, and the vector weight of the parent category is broadcasted to the expanded vector to obtain a comprehensive prediction vector combining the parent category and the subcategory, and the comprehensive prediction vector is used as a label, combined with the document-level feature, to train a classification model; The obtained title or description of the product to be tested is input into the classification model for classification prediction, and a list of related trademarks and patents is recalled as candidate infringement items based on the classification results, and the business scope, NICE and corresponding patent LOC and CPC classification of the trademark matching the product are displayed.

6. A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the product-based intellectual property search and classification method as described in any one of claims 1 to 3.

Images