Trademark protection range discrimination method based on super network model architecture
By adopting a multi-label classification prediction model with a hyper-network model architecture and a label smoothing method in the determination of trademark protection scope, the problem of difficulty in determining trademark protection scope in the prior art is solved, and higher judgment accuracy and robustness are achieved, which is suitable for deployment in low-configuration environments.
Patent Information
- Application Number
- CN202411601205.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-11
- Publication Date
- 2025-03-07
AI Technical Summary
The prior art is difficult to effectively determine the scope of trademark protection in trademark infringement cases, especially when the accuracy of visual description information and insufficient sample size, resulting in low prediction performance and poor robustness.
The trademark protection scope discrimination method based on the hypernetwork model architecture is adopted, and the multi-label classification prediction natural language processing model combined with the label smoothing method is used to analyze the visual similarity characteristics between trademarks through deep learning technology to provide intelligent visual confusion judgment results.
It improves the accuracy and robustness of the determination of trademark protection scope, reduces the model's parameter quantity and computing resource requirements, adapts to the deployment requirements of the low-configuration computing environment, and improves the trial efficiency of trademark infringement cases.
Smart Images
Figure CN119579959A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of natural language processing and machine learning, and relates to a method for determining the protection scope of a trademark based on a hypernetwork model architecture. Background Art
[0002] Intellectual property refers to the ownership of the intellectual labor results produced by human intellectual labor. In recent years, my country has continuously strengthened the protection of intellectual property rights, and has formed institutional guarantee capabilities in intellectual property application, property rights transactions, judicial protection, etc., effectively protecting the legitimate rights and interests of right holders. With the rapid development of my country's market economy, trademarks are no longer just important visual logos of companies or products, but also core intellectual property owned by various organizations and institutions. However, due to the considerable commercial interests they represent and their easy reproducibility, a large number of trademark infringement cases have been generated. According to the statistical information on intellectual property cases released by the Supreme People's Court, the proportion of trademark infringement cases is significantly higher than that of infringement cases of other object categories. The determination of the scope of protection of trademark objects is an important part of the trial of trademark infringement cases, and it has the basis and ability to achieve rapid judgment through technical means, which is of great significance to improving the ability to protect intellectual property rights.
[0003] According to the judicial definition of the scope of protection of trademarks, it can be divided into two categories. First, for non-well-known trademarks, it is prohibited to use the same or similar signs as registered trademarks on the same or similar goods, which is likely to cause confusion. Second, for well-known trademarks, it is prohibited to mix registered well-known trademarks across categories, and unregistered well-known trademarks are prohibited from mixing the same category. According to Article 14 of the newly revised "Trademark Law", manufacturers and operators shall not use the words "well-known trademark" on goods, product packaging or containers, or in advertising, exhibitions and other commercial activities. At the same time, only when the parties request to identify their trademarks as well-known trademarks and the well-known trademarks need to be identified according to the case, the people's court will make a determination in accordance with the law. In the trial of trademark infringement cases, if the accused infringer uses the same or similar trademark as the trademark registrant on the same or similar goods, it is not necessary to determine whether it is a well-known trademark. According to this article, in infringement cases involving well-known trademarks, it is also necessary to judge the confusing nature of the trademarks involved.
[0004] Among them, the confusion judgment standard is the basis for comprehensively judging the distinctiveness of the relevant trademarks, actual usage, whether there is improper intention, the objectively existing market structure, etc. Therefore, the judgment on the scope of trademark protection in the present invention is mainly for the distinctiveness judgment between the trademarks involved. By taking the visual description information between the two trademarks as input to analyze the features including visual similarity features and the similarity of the characters in the trademarks, an intelligent visual confusion judgment result is provided to achieve the judgment of the scope of trademark protection.
[0005] Among the existing related technical solutions, Bhambhoria et al. used deep learning models in the field of trademark law to clarify the possibility of confusion between trademarks in their research. The neural network architecture added a model-independent interpretable intermediate layer, combined with the curriculum learning method to predict the classification of visually confusing labels of trademarks. The purpose of using the curriculum learning mechanism is to achieve weakly supervised learning of the visually confusing description labels of trademarks through this strategy, which can effectively improve the classification performance of deep learning models under few-sample conditions. At the same time, compared with deep neural network models using non-curriculum learning mechanisms, this method has good prediction performance under few-sample conditions and can reduce the cost of sample annotation. The core of this method is to use a curriculum learning mechanism to achieve classification prediction of visually confusing description labels of trademarks under weakly supervised deep learning task conditions. This method requires defining a scheduler for curriculum learning and first distinguishing data samples according to their classification difficulty, which adds additional data processing work costs.
[0006] Hu et al. proposed a solution for the multi-label classification task of trademark confusion in the field of e-commerce on the Chinese Internet. The solution is mainly based on the tuning training method of the pre-trained language model, and obtains a generative description of the classification result by adding classification label placeholders to the generated template. The method was tested on a variety of pre-trained language models, including BERT, ChatGLM, and GPT3.5, and the effectiveness of the method was verified on the provided open source data trademark description dataset. The solution based on the pre-trained language model adopted by this method requires a large amount of training samples, a large number of training iterations, and a high computing environment for model deployment. At the same time, the experiment shows that the prediction accuracy of this method is low under small sample conditions, and the robustness and generalization ability of the model are weak. Summary of the invention
[0007] In view of the problems existing in the prior art, the purpose of the present invention is to provide a trademark protection scope determination method based on a hypernetwork model architecture. This patent uses a hypernetwork as the main architecture, combined with a multi-label classification prediction natural language processing model using a label smoothing method, to achieve the classification prediction goal for the visual confusion descriptions between different trademarks.
[0008] In order to solve the problem of identifying the scope of trademark protection in trademark infringement cases during the process of intellectual property protection, the present invention is based on the intellectual property case dataset, and establishes a multi-label prediction model based on the visual similarity description features of the infringed trademark information based on the case description, to determine whether the trademark involved infringes the protection scope of the protected object, and to determine whether it is an easily confusing trademark based on the labeled output results.
[0009] The present invention uses key information in the case description as input, analyzes and extracts human sensory similarity labels, and annotates human sensory discrimination data labels. The design purpose of this model is to use a deep learning model as a basis, combined with a hypernetwork architecture and a label smoothing mechanism, to enhance the robustness and predictive generalization capabilities of the model, and to solve problems such as low prediction performance caused by the large number of trademark categories and the poor accuracy of trademark visual description information. At the same time, the hypernetwork architecture can effectively reduce the number of parameters of the main model while maintaining prediction accuracy, providing a structural guarantee to meet deployment requirements in low-configuration computing environments. The use of a label smoothing mechanism can effectively improve the prediction accuracy and generalization capabilities of the model in multi-label prediction tasks, so that the model has stronger application capabilities in practical applications.
[0010] The present invention provides a trademark protection scope determination method based on a hypernetwork model architecture, which utilizes natural language understanding and deep learning technology to be applied to the problem of determining the object protection scope of trademark infringement cases. It can effectively improve the multi-label classification prediction efficiency of visual similarity description information between trademarks, which is of great significance to improving the trial efficiency of such cases.
[0011] The present invention provides a multi-label text classification mechanism combining a hypernetwork architecture and label smoothing. In the problem of multi-label text classification, a hypernetwork architecture combined with a label smoothing method is adopted to realize parameter generation and label smoothing processing of an LSTM network, which can effectively reduce the number of parameters of the model and improve the robustness of the model in practical applications.
[0012] The technical solution of the present invention is:
[0013] A method for determining the scope of trademark protection based on a hypernetwork model architecture, the steps of which include:
[0014] 1) Label each case in the trademark infringement case dataset to obtain the training set T train ;
[0015] 2) Construct a multi-label classification model for the protection scope of trademark rights;
[0016] 3) Define the loss function and use the training set T train Performing model training on a multi-label classification model of the protection scope of the trademark right object;
[0017] 4) Inputting the case fact text sequence containing the description information of the trademark involved into the trained multi-label classification model of the protection scope of the trademark right object to obtain the judgment result of the trademark involved.
[0018] Furthermore, the multi-label classification model of the protection scope of the trademark right object includes an encoder module, a label splicing and fusion module, a hypernetwork LSTM module, a decoder LSTM calculation module, a feedforward neural network module, a classifier module, a label smoothing processing module and a divergence loss function calculation module; the method for training the multi-label classification model of the protection scope of the trademark right object is:
[0019] 21) Generate a trademark similarity description text sequence x based on the description data of the kth case k ; x k and the visually confusing class label y for the kth case l Input the encoder module for encoding to obtain the visually confusing category label y l The corresponding text vector L v , trademark similarity description text sequence x k The corresponding trademark description text vector W v ;
[0020] 22) Use the label splicing and fusion module to transform the trademark description text vector W v and the text vector L v Concatenate and merge to form a new encoding vector V fuse , and input them into the hypernetwork LSTM module and the decoder LSTM calculation module respectively;
[0021] 23) The decoder LSTM calculation module calculates the encoding vector V according to the encoding vector V fuse The output result of the hypernetwork LSTM module is decoded as the input of the feedforward neural network, and the result vector is recorded as V forward ;
[0022] 24) V forward As the input of the classifier module, the predicted probability vector P of each label of the kth case is calculated. s ;
[0023] 25) The label smoothing module processes the predicted probability vector P s Perform smoothing calculation to obtain the smoothed label
[0024] 26) Calculate the loss between the smoothed label corresponding to the case and the visual confusion category label annotated with the case, and optimize the multi-label classification model of the protection scope of the trademark right object by minimizing the loss.
[0025] Furthermore, the hypernetwork LSTM module encodes the encoding vector V by the following formula: fuse To process:
[0026] f(t)=σ(dHxf (z Hxf )×W xf x t +d Hhf (z Hhf )×W hf h(t-1)+b f (z fb ))
[0027] i(t)=σ(d Hxi (z Hxi )×W xi x t +d Hhi (z Hhi )×W hi h(t-1)+b i (z ib ))
[0028]
[0029] o(t)=σ(d Hxo (z Hxo )×W xo x t +d Hho (z Hho )×W ho h(t-1)+b o (z ob ))
[0030] h(t)=o(t)×tanh(c(t))
[0031] Among them, x t is the input data at time t, W xf is the computable parameter matrix of the forget gate, W hf Calculate the parameter matrix for the previous state hidden layer of the forget gate; d Hxf (z Hxf ) is one of the calculation outputs of the hypernetwork LSTM calculation module;
[0032]
[0033] in is one of the outputs of the hypernetwork LSTM computation module, for The output gate trainable parameter matrix, Calculate the bias for the output gate, W xfz Output z Hxf The trainable parameter matrix of Hhf (z Hhf ) is calculated as:
[0034]
[0035] W xi Compute the parameter matrix for the input gate, W hi Calculate the parameter matrix for the hidden layer of the previous state of the input gate; d Hxi (z Hxi ) is one of the input gate calculation outputs of the hypernetwork LSTM calculation module;
[0036]
[0037] for The previous state in the input gate of is a trainable parameter matrix, Calculate the bias for the input gate calculation result, W xiz Calculate the result z for the input gate Hxi The trainable parameter matrix of Hhi (z Hhi ) is calculated as:
[0038]
[0039] W xc Calculate the parameter matrix for the cell state, W hc Calculate the parameter matrix of the previous state hidden layer in the cell state calculation formula; d Hxc (z Hxc ) is one of the calculation outputs of the hypernetwork LSTM calculation module;
[0040]
[0041] The previous cell state The trainable parameter matrix, Bias in the cell state calculation formula, W xcz Calculate the result z for the cell state Hxc The trainable parameter matrix of Hhc (z Hhc ) is calculated as:
[0042]
[0043] W xo Compute the parameter matrix for the output gate, W ho Calculate the parameter matrix for the previous state hidden layer of the output gate calculation result;
[0044] d Hxo (z Hxo ) is one of the calculation outputs of the hypernetwork LSTM calculation module;
[0045]
[0046] The previous state of the output gate The trainable parameter matrix, The bias in the output gate calculation formula, W xoz Calculate the result z for the output gate Hxo The trainable parameter matrix of Hho (z Hho ) is calculated as:
[0047]
[0048] Further, the training set T is obtained train The method is:
[0049] 11) By searching and sorting out the civil and criminal cases involving trademarks disclosed on the Judgment Documents Network, a data set T is formed. cases ;
[0050] 12) According to the label set Y, the dataset T cases The case description data in the label set Y are labeled; the labels in the label set Y include the following labels: similar patterns, identical patterns, similar texts, identical texts, identical trademarks, and different patterns and texts.
[0051] Furthermore, the label smoothing module uses the formula For the predicted probability vector P s Perform smoothing calculation to obtain the smoothed label result Where L is the number of labels, if the predicted probability vector P s The label y in l 'If y exists in the label set Y l ' is 1 if yes, otherwise it is 0.
[0052] Furthermore, the classifier module is a classifier softmax function.
[0053] Furthermore, the encoder module is a BERT pre-trained language model.
[0054] Furthermore, the cross entropy loss function is used to calculate the loss between the smoothed label corresponding to the case and the visually confusing category label annotated with the case.
[0055] A server, characterized in that it comprises a memory and a processor, wherein the memory stores a computer program, the computer program is configured to be executed by the processor, and the computer program comprises instructions for executing the above method.
[0056] A computer-readable storage medium stores a computer program, wherein the computer program implements the above method when executed by a processor.
[0057] The advantages of the present invention are as follows:
[0058] This method uses a deep neural network method based on label smoothing and a hypernetwork model architecture to predict the scope of object protection in trademark infringement cases. The use of a new network structure combined with a label smoothing processing mechanism can effectively reduce the overfitting of the model and improve its robustness in actual predictions. At the same time, the hypernetwork-based model architecture can effectively reduce the number of parameters used in the model, which is of great significance for reducing the model's disk usage and accelerating the prediction speed. BRIEF DESCRIPTION OF THE DRAWINGS
[0059] Figure 1 This is the model architecture diagram in this method.
[0060] Figure 2 Flow chart of this method. DETAILED DESCRIPTION
[0061] The present invention is further described in detail below in conjunction with the accompanying drawings. The examples given are only used to explain the present invention but not to limit the scope of the present invention.
[0062] The present invention proposes a trademark protection scope determination method based on a hypernetwork model architecture, which is used to determine whether the trademark object involved in the case violates the legal provisions on the trademark protection scope in my country's Trademark Law. In this invention, a method based on a hypernetwork model architecture combined with label smoothing is adopted to improve the robustness and prediction accuracy of multi-label classification prediction problems in trademark protection scope prediction. The following will be combined with the accompanying drawings in the examples of the present invention to clearly and completely describe the technical solution in the examples of the present invention. Figure 1 It is a flow chart of the technical method implemented by the present invention. Figure 2 It is a model structure diagram implemented by the present invention. The technical method includes the following steps:
[0063] (1) Generate a trademark infringement case dataset;
[0064] (2) Construct a multi-label classification model for the scope of protection of trademark rights;
[0065] (3) Define the loss function and perform model training;
[0066] (4) By inputting a case fact text sequence containing descriptive information of the trademark involved, a multi-label determination result on the protection scope of the trademark involved is returned.
[0067] Step 1 includes the following sub-steps:
[0068] Step 1.1: Retrieve and sort out the civil and criminal cases involving trademarks published on the Judgment Documents Network to form a dataset T for model training and testing, which includes the IP civil case dataset T civial And the criminal case dataset T criminal See formula (1).
[0069] T=T civial ∪T criminal (1)
[0070] Step 1.2: Screen and filter out cases that the court has no jurisdiction over or has not filed. See formula (2), where T non-jurisdiction For cases without jurisdiction, T unregistered This is an unopened case.
[0071] T cases =TT non-jurisdiction -T unregistered (2)
[0072] Step 1.3: According to the provisions of the Trademark Law, it is prohibited to use the same or similar marks as the registered trademark on the same or similar goods, which may easily lead to confusing similarity. Based on this legal requirement, the problem of identifying the scope of protection of trademarks is modeled as a multi-label classification problem of visual discrimination description between the trademarks involved. The dataset T is manually annotated. cases The case description data in is annotated with a single label, and the real value label set Y is defined as follows:
[0073] Y = {y1 = similar patterns, y2 = identical patterns, y3 = similar texts, y4 = identical texts, y5 = identical trademarks, y6 = different graphics and texts} (3)
[0074] Step 1.4: Split the dataset into training set T according to the different purposes of training and testing train , validation set T val And the test set T test :
[0075] T cases =T train ∪T val ∪T test (4)
[0076] Step 2 includes the following sub-steps:
[0077] Step 2.1: Define the multi-label classification problem of trademark right object protection scope as f(x k ; θ) = y k , where x kDefined as a text sequence with the case description or the visual similarity description of the trademark involved as the main content As the input of the model, {k1,k2,k3....k n} is the sequential number of the elements in the sequence. θ is defined as the parameters contained in the model. The training process is to minimize the loss function by optimizing θ, that is:
[0078]
[0079] Where δ is the loss function used by the model to evaluate the model's predicted label y k And the data true value label y l The loss between them, N is the number of a batch of data during the training process.
[0080] Step 2.2, the model is mainly divided into eight components, namely encoder module, label splicing and fusion module, hypernetwork LSTM module, decoder LSTM calculation module, feedforward neural network module, classifier module, label smoothing processing module and divergence loss function calculation module. The encoder module uses the pre-trained language model BERT as the text embedding encoder.
[0081] Step 2.3: During the training process, the BERT pre-trained language model is first used as the encoder module to convert the trademark similarity description text sequences into Visually confusing category label y l Encoded into a text vector. Define the trademark description text vector as W v , the encoding vector of the visually confusing label group is defined as L v .
[0082] Step 2.4: In the label concatenation and fusion module, the text vector W v and visually confusing label group L v As input, vector concatenation is used to form a new encoding vector V fuse .
[0083] Step 2.5: vector V fuse As the input of the hypernetwork LSTM module and the decoder LSTM calculation module, the output of the hypernetwork LSTM will be used as the decoder LSTM calculation module parameter. The calculation formulas of each gate in the decoder LSTM calculation module are as follows:
[0084] f(t)=σ(d Hxf (z Hxf )×W xf x t +d Hhf (z Hhf )×W hf h(t-1)+bf (z fb )) (6)
[0085] i(t)=σ(d Hxi (z Hxi )×W xi x t +d Hhi (z Hhi )×W hi h(t-1)+b i (z ib )) (7)
[0086]
[0087] o(t)=σ(d Hxo (z Hxo )×W xo x t +d Hho (z Hho )×W ho h(t-1)+b o (z ob )) (10)
[0088] h(t)=o(t)×tanh(c(t)) (11)
[0089] In the above formula, x t is the input at time t, and formula (6) is the calculation formula of the forget gate. xf is the computable parameter matrix of the forget gate, W hf Calculate the parameter matrix for the previous state hidden layer of the forget gate. Hxf (z Hxf ) is the calculation output of the hypernetwork LSTM calculation module, as shown below:
[0090]
[0091] in is the output of the hypernetwork LSTM computing module, for The output gate trainable parameter matrix, Calculate the bias for the output gate, W xfz Output z Hxf The trainable parameter matrix of . Similar to the calculation formula, d Hhf (z Hhf ) is calculated as follows:
[0092]
[0093] Formula (7) is the input gate calculation formula where W xiCompute the parameter matrix for the input gate, W hi Calculate the parameter matrix for the previous state hidden layer of the input gate. Hxi (z Hxi ) is the input gate calculation output of the hypernetwork LSTM calculation module, as shown below:
[0094]
[0095] for The previous state in the input gate of is a trainable parameter matrix, Calculate the bias for the input gate calculation result, W xiz Calculate the result z for the input gate Hxi The trainable parameter matrix of . Similar to the calculation formula, d Hhi (z Hhi ) is calculated as follows:
[0096]
[0097] Formula (8) is the cell state calculation formula where W xc Calculate the parameter matrix for the cell state, W hc Calculate the parameter matrix of the previous state hidden layer in the cell state calculation formula. Hxc (z Hxc ) is the calculation output of the hypernetwork LSTM calculation module, as shown below:
[0098]
[0099] The previous cell state The trainable parameter matrix, Bias in the cell state calculation formula, W xcz Calculate the result z for the cell state Hxc The trainable parameter matrix of . Similar to the calculation formula, d Hhc (z Hhc ) is calculated as follows:
[0100]
[0101] Formula (9) is the update formula of the memory unit, which realizes the update of the memory cell.
[0102] Formula (10) is the output gate calculation formula where W xo Compute the parameter matrix for the output gate, W ho Calculate the parameter matrix for the previous state hidden layer of the output gate calculation result. Hxo (z Hxo ) is the calculation output of the hypernetwork LSTM calculation module, as shown below:
[0103]
[0104] The previous state of the output gate The trainable parameter matrix, The bias in the output gate calculation formula, W xoz Calculate the result z for the output gate Hxo The trainable parameter matrix of . Similar to the calculation formula, d Hho (z Hho ) is calculated as follows:
[0105]
[0106] Formula (11) is the formula for calculating the hidden layer state at time t.
[0107] Step 2.6: After the decoder LSTM calculation module calculates, the decoder calculation result is used as the input of the feedforward neural network, and the result vector is recorded as V forward .
[0108] Step 2.7: V forward As the input of the classifier softmax function, the predicted probability vector P of each label is calculated s .
[0109] P s =softmax(V forward ) (20)
[0110] Step 3 includes the following sub-steps:
[0111] Step 3.1: Take the label as the input of the label smoothing module and use the following formula to complete the smoothing calculation of the data label, where L is the number of labels. l 'If y exists in the label set Y corresponding to the input l is 1, otherwise it is 0. is the label result after smoothing.
[0112]
[0113] Step 3.2: Calculate the loss of the model after label smoothing through the cross entropy loss function as shown in formula (22), where the subscript i refers to the data sequence number in a batch of training data, l refers to the sequence number of the label in the label set Y, is the smoothed input data x i The annotated true value label, p il is the probability of predicting the true value label.
[0114]
[0115] Step 4 includes the following sub-steps:
[0116] Step 4.1: The user inputs the case description text, and the model outputs the predicted visual description similarity label set between the two trademarks.
[0117] Step 4.2: Use the output visual description similarity label to determine whether the suspected infringing trademark infringes the protection scope of the infringed trademark object. That is, if the output label is y1~y5, it infringes the protection scope of the trademark object. If the label is y6, it can be determined that there is no infringement.
[0118] Although the specific embodiments of the present invention are disclosed for the purpose of illustration, the purpose is to help understand the content of the present invention and implement it accordingly, those skilled in the art will understand that various substitutions, changes and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. Therefore, the present invention should not be limited to the content disclosed in the best embodiment, and the scope of the present invention is subject to the scope defined in the claims.
Claims
1. A method for determining the scope of trademark protection based on a hypernetwork model architecture, the steps of which include: 1) Label each case in the trademark infringement case dataset to obtain the training set T train ; 2) Construct a multi-label classification model for the protection scope of trademark rights; 3) Define the loss function and use the training set T train Performing model training on a multi-label classification model of the protection scope of the trademark right object; 4) Inputting the case fact text sequence containing the description information of the trademark involved into the trained multi-label classification model of the protection scope of the trademark right object to obtain the judgment result of the trademark involved.
2. The method according to claim 1, characterized in that The multi-label classification model for the protection scope of the trademark right object includes an encoder module, a label splicing and fusion module, a hypernetwork LSTM module, a decoder LSTM calculation module, a feedforward neural network module, a classifier module, a label smoothing processing module and a divergence loss function calculation module; the method for training the multi-label classification model for the protection scope of the trademark right object is: 21) Generate a trademark similarity description text sequence x based on the description data of the kth case k ; x k and the visually confusing class label y for the kth case l Input the encoder module for encoding to obtain the visually confusing category label y l The corresponding text vector L v , trademark similarity description text sequence x k The corresponding trademark description text vector W v ; 22) Use the label splicing and fusion module to transform the trademark description text vector W v and the text vector L v Concatenate and merge to form a new encoding vector V fuse , and input them into the hypernetwork LSTM module and the decoder LSTM calculation module respectively; 23) The decoder LSTM calculation module calculates the encoding vector V according to the encoding vector V fuse The output result of the hypernetwork LSTM module is decoded as the input of the feedforward neural network, and the result vector is recorded as V forward ; 24) V forward As the input of the classifier module, the predicted probability vector P of each label of the kth case is calculated. s ; 25) The label smoothing module processes the predicted probability vector P s Perform smoothing calculation to obtain the smoothed label 26) Calculate the loss between the smoothed label corresponding to the case and the visual confusion category label annotated with the case, and optimize the multi-label classification model of the protection scope of the trademark right object by minimizing the loss.
3. The method according to claim 2, characterized in that The hypernetwork LSTM module encodes the encoding vector V by the following formula fuse To process: f(t)=σ(d Hxf (z Hxf )×W xf x t +d Hhf (z Hhf )×W hf h(t-1)+b f (z fb )) i(t)=σ(d Hxi (z Hxi )×W xi x t +d Hhi (z Hhi )×W hi h(t-1)+b i (z ib )) o(t)=σ(d Hxo (z Hxo )×W xo x t +d Hho (z Hho )×W ho h(t-1)+b o (z ob )) h(t)=o(t)×tanh(c(t)) Among them, x t is the input data at time t, W xf is the computable parameter matrix of the forget gate, W hf Calculate the parameter matrix for the previous state hidden layer of the forget gate; d Hxf (z Hxf ) is one of the calculation outputs of the hypernetwork LSTM calculation module; in is one of the outputs of the hypernetwork LSTM computation module, for The output gate trainable parameter matrix, Calculate the bias for the output gate, W xfz Output z Hxf The trainable parameter matrix of Hhf (z Hhf ) is calculated as: W xi Compute the parameter matrix for the input gate, W hi Calculate the parameter matrix for the hidden layer of the previous state of the input gate; d Hxi (z Hxi ) is one of the input gate calculation outputs of the hypernetwork LSTM calculation module; for The previous state in the input gate of is a trainable parameter matrix, Calculate the bias for the input gate calculation result, W xiz Calculate the result z for the input gate Hxi The trainable parameter matrix of Hhi (z Hhi ) is calculated as: W xc Calculate the parameter matrix for the cell state, W hc Calculate the parameter matrix of the previous state hidden layer in the cell state calculation formula; d Hxc (z Hxc ) is one of the calculation outputs of the hypernetwork LSTM calculation module; The previous cell state The trainable parameter matrix, Bias in the cell state calculation formula, W xcz Calculate the result z for the cell state Hxc The trainable parameter matrix of Hhc (z Hhc ) is calculated as: W xo Compute the parameter matrix for the output gate, W ho Calculate the parameter matrix of the previous hidden layer of the output gate calculation result; d Hxo (z Hxo ) is one of the calculation outputs of the hypernetwork LSTM calculation module; The previous state of the output gate The trainable parameter matrix, The bias in the output gate calculation formula, W xoz Calculate the result z for the output gate Hxo The trainable parameter matrix of Hho (z Hho ) is calculated as:
4. The method according to claim 1, 2 or 3, characterized in that: Get the training set T train The method is: 11) By searching and sorting out the civil and criminal cases involving trademarks published on the Judgment Documents Network, a data set T is formed. cases ; 12) According to the label set Y, the dataset T cases The case description data in the label set Y are labeled; the labels in the label set Y include the following labels: similar patterns, identical patterns, similar texts, identical texts, identical trademarks, and different patterns and texts.
5. The method according to claim 4, characterized in that The label smoothing module uses the formula For the predicted probability vector P s Perform smoothing calculation to obtain the smoothed label result Where L is the number of labels, if the predicted probability vector P s The label y in l 'If y exists in the label set Y l ' is 1 if yes, otherwise it is 0.
6. The method according to claim 1, 2 or 3, characterized in that: The classifier module is a classifier softmax function.
7. The method according to claim 1, 2 or 3, characterized in that: The encoder module is a BERT pre-trained language model.
8. The method according to claim 1, characterized in that The cross entropy loss function is used to calculate the loss between the smoothed label corresponding to the case and the visually confusing category label annotated with the case.
9. A server, characterized in that: The invention comprises a memory and a processor, wherein the memory stores a computer program, the computer program is configured to be executed by the processor, and the computer program comprises instructions for executing the method according to any one of claims 1 to 8.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 8 is implemented.