Method for identifying protection range of copyright case object

Abstract

The invention discloses a method for identifying a copyright case object protection range. The method comprises the following steps: 1) acquiring and marking a plurality of copyright infringement cases to obtain a copyright infringement case data set; the annotation information comprises the field range of the works, the object form of the works and the originality of the works; 2) establishing a copyright object protection range judgment model for predicting the field range of the works, the object form and the work originality; the copyright object protection range judgment model comprises an encoder module, a joint attention calculation module, a splicer module, a multi-layer perceptron module and a classifier module; 3) using the copyright infringement case data set to train a copyright object protection range determination model; and 4) inputting a to-be-identified work into the trained copyright object protection range judgment model to obtain a field range, an object form and originality of the work. According to the method, intelligent judgment on the protected range of the copyright object can be realized through the case description information.

Description

Technical Field

[0001] The present invention belongs to the field of natural language processing and machine learning technology, and relates to a method for identifying the scope of protection of an object in a copyright case. Background Art

[0002] my country attaches great importance to the protection of intellectual property rights, particularly in the area of ​​judicial protection. People's courts have continuously strengthened their adjudication of works, trademarks, and patents, enhancing incentives for cultural, artistic, and technological innovation, and promoting technological innovation and industrial upgrading. However, with the rapid development of my country's economy, society, and science and technology, the number of intellectual property infringement cases has been rising year by year. People's courts at all levels are faced with practical challenges, including a large number of cases and a small number of judges, the wide range of intellectual property subject matter involved, and significant technical complexity. A study of publicly available judicial documents in intellectual property cases reveals that determining the scope of protection for intellectual property subject matter is a fundamental step in the trial process and a crucial factor in determining the outcome of a case. This is particularly true in copyright cases, where the law clearly defines a wide range of categories of protected works, making the determination of the scope of protection somewhat challenging.

[0003] At the same time, new information technologies, represented by machine learning, particularly deep neural networks, have achieved promising results in natural language processing, computer vision, and big data analysis. Research on the application of intelligent judicial assistance has yielded significant breakthroughs in areas such as crime prediction, applicable laws and regulations, judicial document summaries, and judicial question-and-answering, through the study of judicial documents and the application of relevant models. However, specialized research on intellectual property cases has primarily focused on the prediction and determination of trademark similarity, design patent similarity, and literary work similarity, lacking intelligent assistance technologies specifically tailored to the elements of intellectual property case adjudication. Unlike trademarks and patents, copyrights are more complex in defining the scope of protection, including originality, artistry, the scope of the subject matter, and the form of the work. Therefore, technological approaches are more necessary to provide judges with predictive advice, which is also more valuable in judicial practice.

[0004] This patent addresses the issue of determining the scope of protection of copyright objects in copyright cases and proposes a multi-category prediction method based on label embedding to achieve intelligent determination of the scope of protection of copyright objects through case description information.

[0005] In an existing related technical solution, Wang et al. transformed the text classification problem into a joint embedding problem between the target text information and the classification labels, using label embedding to improve the accuracy of multi-label prediction (see G. Wang et al., "Joint Embedding of Words and Labels for Text Classification." arXiv, May 10, 2018. Accessed: July 21, 2023. [Online]. Available: http: / / arxiv.org / abs / 1805.04174). In this solution, the text and labels are first mapped into the same word embedding space. The text word vectors are then fused with the label word vectors by calculating cosine. Finally, a single-layer perceptron and a softmax function are used to calculate the compatibility score to complete the category and label classification task. This method uses the 300-dimensional GloVe word embedding model to calculate the word vectors for the text and labels, and primarily uses cosine calculations and a single-layer perceptron to calculate the compatibility score. The essence of the algorithm is to obtain the attention of the label to the input text. The attention effect obtained by this method still lags behind the currently used multi-head attention calculation method.

[0006] Tan et al. proposed a generative network that adaptively generates convolutional filters. This network uses the generated convolutional neural network to classify word embeddings fused with text and labels (see C. Tan, Y. Ren, and C. Wang, “Anadaptive convolution with label embedding for text classification,” ApplIntell, vol. 53, no. 1, pp. 804–812, Jan. 2023, doi:10.1007 / s10489-021-02702-x). This approach uses the BERT pre-trained model to calculate word embeddings for text and labels. The generative network for adaptive convolutional filters is constructed using a Bi-GRU, an attention mechanism, and a fully connected network. Classification results are then obtained through computations in the convolutional neural network layer and the classification function. The generative network of the adaptive generative convolution filter adopted in this method can effectively fill the convolution blocks in subsequent calculations to achieve adaptive convolution. However, when processing datasets with a large number of categories and high similarity between categories, its performance is not much improved compared to the accuracy of the baseline model, and it cannot handle multi-label and multi-task classification problems. Summary of the Invention

[0007] In response to the problems existing in the prior art, the purpose of the present invention is to provide a method for identifying the scope of protection of the object of copyright case, which can realize intelligent determination of the scope of protection of the copyright object through case description information.

[0008] In order to solve the problem of difficulty in determining the scope of protection of the objects involved in copyright infringement cases, according to Article 3 of the Copyright Law of the People's Republic of China, "The works referred to in this Law refer to intellectual achievements in the fields of literature, art and science that are original and can be expressed in a certain form, including: (1) written works; (2) oral works; (3) music, drama, folk art, dance, and acrobatic art works; (4) fine arts and architectural works; (5) photographic works; (6) audiovisual works; (7) engineering design drawings, product design drawings, maps, schematic diagrams and other graphic works and model works; (8) computer software; (9) other intellectual achievements that meet the characteristics of works." The determination of the scope of protection of the objects involved in the case is converted into a determination of the scope of the field to which they belong, the form of the work and the uniqueness of the work.

[0009] This problem involves multi-label, multi-task natural language classification of text describing case facts. Existing technical approaches primarily employ specialized models trained using pre-trained language models, computational attention between labels and input text, and the design of multi-task frameworks. However, there is currently a lack of technical solutions for determining the scope of protection for subject matter in copyright infringement cases. Furthermore, due to the correlation between the three label classification issues—scope of the field, form of the work, and uniqueness of the work—the impact of the scope of the field and form on the uniqueness of the work must be considered during the determination process.

[0010] Therefore, it is necessary to design a model structure with the ability to process multiple classification tasks to address this problem. It can infer the scope of the subject matter involved, the form of the work, and the uniqueness of the work from the case description information, and can calculate the mutual influence between these three types of sub-tasks.

[0011] At the same time, multi-layer joint attention neural networks lead to high algorithmic complexity and excessively long training and prediction times, making them unsuitable for online applications. Therefore, this method innovatively combines a joint attention mechanism with a multi-layer neural network model based on the agent attention mechanism originally applied in image processing (see D. Han et al., “AgentAttention: On the Integration of Softmax and Linear Attention,” arXiv, arXiv:2312.08874, Jul. 2024. doi:10.48550 / arXiv.2312.08874). This reduces algorithmic complexity while improving the ability to capture global features of long texts, leveraging the label fusion model training mechanism for multi-task text classification and prediction scenarios.

[0012] Based on the specific trial scenario requirements of the model application, the fields to which the objects that may be involved in intellectual property cases belong are extremely broad and are constantly changing with the development of technological innovation and artistic creation. Therefore, the judgment model for the scope of protection of intellectual property objects should also have a certain generalization prediction ability and adaptability for predicting data and fields. In the invention of this method, a domain generalization technology method based on test-time adaptation is incorporated (reference D.Wang, E.Shelhamer, S.Liu, B.Olshausen, and T.Darrell, “Tent: Fully Test-time Adaptation by EntropyMinimization,” Mar.18, 2021, arXiv:arXiv:2006.10726.doi:10.48550 / arXiv.2006.10726), which enables the model to adapt to changes in the statistical distribution of data during online service.

[0013] This paper provides a method for predicting the scope of protection of subject matter in copyright infringement cases. Using a multi-task deep learning framework for natural language understanding, it identifies the scope of the subject matter, its form, and the uniqueness of the work. It also employs a test-time adaptive approach to improve the model's ability to adapt to data generalization during testing and prediction.

[0014] The present invention provides a hybrid attention mechanism for label embedding under a multi-task model framework. It combines the joint attention and proxy attention of the application text and the label embedding to be predicted under the multi-task text classification model framework, improving the accuracy of each subtask prediction while reducing the overfitting risk of the subtask.

[0015] The technical solution of the present invention is:

[0016] A method for identifying the scope of protection of an object in a copyright case, comprising the following steps:

[0017] 1) Obtain multiple copyright infringement cases and annotate them to obtain a copyright infringement case dataset; the annotated information includes the field to which the work belongs, the object form of the work, and the originality of the work;

[0018] 2) establishing a copyright object protection scope determination model to predict the scope of the work, the object form, and the originality of the work; the copyright object protection scope determination model includes an encoder module, a joint attention calculation module, a splicer module, a multi-layer perceptron module, and a classifier module;

[0019] 3) Using the copyright infringement case dataset to train the copyright object protection scope determination model; wherein, during the training process, the encoder module encodes the input work description text W and its annotation information to obtain a text vector, the text vector including the work description text vector W n , the object's domain category label vector L dv , object form category label vector L fv And the work originality category label vector L sv ; The joint attention calculation module calculates W v With L dv The multi-head attention vector between W v With L fv The multi-head attention vector between and W v With L sv The multi-head attention vector W between att|Ls ; The splicer module combines the three multi-head attention vectors W att|Ls Splice to form a vector W concate ; The joint attention calculation module converts the vector W concate Respectively with the query weight matrix W Q , key weight matrix W K , value weight matrix W V Multiply and calculate Q A =W Q W concate , K A =W K W concate 、V A =W V W concate ; The multi-layer perceptron module mlp A Q AThe proxy token A is calculated as input; the encoder module uses the relative position encoding calculation method to encode the work description text W to obtain the encoding vector PE; the joint attention calculation module is based on the encoding vector PE, proxy token A, Q A , K A 、V A The agent attention value O is calculated A , and then the agent attention value O A With vector W concate Add up to get the residual calculation result and will The output result obtained by inputting the multilayer perceptron module is Add together to get O AE ; Then calculate the aggregation result S L =softmax(mlp aggr (O AE ))×O AE , the classifier module is based on the aggregation result S L Predict the work description text W to which the work belongs, and the prediction result logit domain , Work object form prediction result logit form And the prediction result logit of the originality of the work peculiarity ; Then, the loss value is calculated based on the prediction results and the annotation results to optimize the copyright object protection scope determination model;

[0020] 4) Inputting the work to be identified into the trained copyright object protection scope determination model to obtain the domain scope, object form and originality of the work.

[0021] Furthermore, the multi-layer perceptron module mlp is used A , with Q A The proxy token A is calculated for the input; among them, the multi-layer perceptron module Linear is a linear function and ReLU is an activation function.

[0022] Furthermore, the proxy attention calculation formula O is used A =Attention(Q A A T +PE)Attention(AK AT +PE)V A +conv(V A ) Calculate the agent attention value O A , where conv is a one-dimensional convolutional neural network.

[0023] Furthermore, the cross entropy loss function is used to predict the result logit according to the scope of the work.domain The loss value L is calculated based on the scope of the field to which the annotated work belongs. domin , using the cross entropy loss function to predict the result logit according to the object form of the work form The loss value L is calculated based on the object form of the annotated work form , using the cross entropy loss function to predict the result logit based on the originality of the work peculiarity The loss value L is calculated based on the originality of the marked work unique ; Then calculate the total loss value L = αL domin +βL form +γL unique , α+β+γ=1, and the copyright object protection scope determination model is optimized according to the total loss value L.

[0024] Furthermore, in step 4), the parameter vectors λ and η are defined; the domain range p'(y fc ), object form p'(y dc ) and originality p'(y pc ) to perform linear operations Then calculate the prediction entropy Then calculate the partial derivatives of H(y) with respect to λ and η respectively The updated λ and η are used for the next prediction.

[0025] Furthermore, the copyright infringement cases include civil copyright infringement cases and criminal copyright infringement cases.

[0026] Furthermore, the encoder module uses a pre-trained language model as a text embedding encoder.

[0027] A server, characterized in that it includes a memory and a processor, the memory stores a computer program, the computer program is configured to be executed by the processor, and the computer program includes instructions for executing the above method.

[0028] A computer-readable storage medium stores a computer program thereon, wherein the computer program implements the above method when executed by a processor.

[0029] The advantages of the present invention are as follows:

[0030] This method uses a multi-task text classification method based on deep neural networks, combined with a joint attention mechanism, to predict the scope of protection of the subject matter involved in copyright cases. The multi-task model is trained and implemented to discriminate the scope, form, and uniqueness of the subject matter involved. Because each subtask uses the same contextual semantic encoding layer, the semantics of the case facts described in each subtask are unified, improving the prediction accuracy of each subtask.

[0031] First, by adopting a multi-task mechanism in the method, the model reduces the degree of overfitting of a single task during the learning process, so that the model achieves a regularization effect as a whole, thereby improving the model's generalization ability and versatility in tasks.

[0032] The attention value between the classification label and the case description text is calculated through the joint attention mechanism and the proxy attention mechanism to obtain the correlation information between the case description and the embedded label. The label information is used to construct a representation with text-to-label attention, which improves the prediction accuracy for the three subtasks.

[0033] The test-time adaptive method is applied in the prediction stage to improve the overall confidence of the prediction task and enhance the data generalization ability of the model in the long-term online prediction process. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0035] The present invention will be described in further detail below with reference to the accompanying drawings. The examples given are only used to explain the present invention and are not used to limit the scope of the present invention.

[0036] The present invention proposes a deep neural network method for determining the protection scope of works. Figure 1 Flowchart of the technical method implemented by the present invention, which includes the following steps:

[0037] (1) Processing the copyright infringement case dataset, constructing a dataset that includes civil and criminal copyright infringement cases, defining three prediction subtasks for the scope of the work, the object form of the work, and the originality, and completing data annotation;

[0038] (2) Establish a model for determining the scope of copyright protection, so that the model can simultaneously predict the scope of the field to which the work belongs, the form of the object, and the originality;

[0039] (3) Define the model combination loss function to balance the subtasks of model training;

[0040] (4) After multiple iterations on the training set, the model predicts the scope of the work's field, the form of the work's object, and the degree of originality based on the input case facts or object description information. In order to enhance the model's ability to adapt to data generalization during online prediction, a test-time adaptive method is applied to the model's prediction process. This method reduces the information entropy of the model's prediction results, thereby improving the model's prediction confidence and accuracy, and making it adaptable to changes in the statistical distribution of the data.

[0041] Step 1 includes the following sub-steps:

[0042] Step 1.1: The basic data set is the intellectual property cases published on the Judgment Documents Network, including civil and criminal cases. civil is a civil case dataset, S criminal is a criminal case dataset, S cases For the complete data set.

[0043] S cases =S civil ∪S criminal (1)

[0044] Step 1.2: Screen and filter out cases where the court has no jurisdiction or has not filed a case. non-jurisdiction is a dataset of cases where the court has no jurisdiction, S unregistered This is an unregistered data set.

[0045] S cases =S cases -S non-jurisdiction -S unregistered (2)

[0046] Step 1.3: According to the legal provisions of the Copyright Law, the prediction task is defined as including three subtasks, namely, the classification and judgment of the scope of the work, the object form of the work, and the originality of the work. protection-range Defined as the overall task of protection range prediction, Task domain For domain-wide prediction subtasks, Task form For the object form prediction subtask, Task unique For the originality prediction subtask.

[0047] Task protection-range =Task domain +Task form +Task unique (3)

[0048] And the classification task is modeled as:

[0049]

[0050] Where F is the prediction model, and the case description text sequence is defined as w1,w1,...,w n As the input of the model, the prediction output results of the three subtasks are obtained through the calculation of the model, that is, the prediction results of the domain range Prediction of the work's form and the prediction results of originality

[0051] Step 1.3: Use manual annotation methods to mark the scope of the field, form and uniqueness of the works involved.

[0052] Step 2 includes the following sub-steps:

[0053] Step 2.1: The model consists of five main components: encoder module, joint attention calculation module, splicer module, multi-layer perceptron module, and classifier module. The encoder module uses a pre-trained language model as the text embedding encoder.

[0054] Step 2.2: The case description can be defined as a string of length n W = {w1,w2,w3,..,w n}, and the domain category of the object is defined as a text label array of length i The subscript d above identifies the label of the domain. The object form category is defined as a text label array of length j. The superscript f is used to identify the object form label. The originality category is defined as The superscript S above indicates the originality of the label. Except for the uniqueness category, which is defined as a binary category label, all other categories are multi-category labels.

[0055] Step 2.3: First, use the pre-trained language model as the encoder module to encode the case description text, the object domain category, the object form category, and the uniqueness category into text vectors. All the word vectors after the conversion have the same dimension. Define the case description text vector as W n , the object domain category label vector is defined as L dv , the object form category label vector is defined as L fv And the unique class label vector is defined as L sv .

[0056] Step 2.4: In the joint attention calculation module, refer to the multi-head joint attention calculation method adopted by Liu et al. (reference M.Liu, L.Liu, J.Cao, and Q.Du, "Co-attention network with label embedding for text classification," Neurocomputing, vol. 471, pp. 61-69, Jan. 2022, doi: 10.1016 / j.neucom.2021.10.099). First, define the attention calculation formula as:

[0057]

[0058] In the formula Compute parameters for attention queries, Compute parameters for the attention key, Calculate the parameters for the attention value, The square root of the dimension of the parameter V is calculated for the attention value, and the calculation formula of the multi-head attention is defined as:

[0059] MultiHead(Q,K,V)=softmax([H1;...H p ])W o (6)

[0060] In this formula, W o is the output weight trainable matrix, multi-head attention H i It is calculated by the following attention calculation formula.

[0061] H i =Attention(W Q W n ,W K W n ,W V W n ) (7)

[0062] Where W Q 、W K 、W V is the trainable parameter matrix, where W Q is the query weight matrix, W K is the key weight matrix, W V is the value weight matrix. Using the calculation method of formula (5), W v 、L dv 、L fv 、L sv Substitute them into the formula to calculate the multi-head attention vector between the text and each classification label. The calculation formula is as follows.

[0063]

[0064] in is the multi-head attention vector between the case description text vector and the object domain classification vector, is the multi-head attention vector between the case description text vector and the object form classification vector, W att|Ls It is the multi-head attention vector between the case description text vector and the uniqueness classification vector.

[0065] Step 2.5: Input the three attention vectors into the splicer module and form a vector W by vector splicing. concate ,

[0066] Step 2.6: Multiply the input by the three trainable parameter matrices W Q 、W K 、W V And get the query Q used in the attention calculation A , key K A Sum V A .

[0067]

[0068] Step 2.7, using the multi-layer perceptron model mlp A , with Q A The proxy token A is calculated for the input. In this step, the multi-layer perceptron mlp A The definition is shown in formula (10), which is defined by two layers of linear function Linear plus one layer of activation function ReLU, where Q L and Q A is the intermediate calculation result.

[0069]

[0070] Formula (11) is obtained by mlp A The calculation formula for the proxy token A is calculated.

[0071] A=mlp A (Q) (11)

[0072] Step 2.8: Use relative position encoding calculation method to calculate the input W = {w1,w2,w3,..,w n}Calculate the encoding vector PE for each position, where i is the position parameter in the input and d is the dimension value of the input sequence length.

[0073]

[0074] Step 2.9: Use the proxy attention calculation formula (13) to calculate the proxy attention value O A , where conv is a one-dimensional convolutional neural network, with V A As input to increase the diversity of calculation results. , where A T is the transpose of A, K AT K A The transpose of .

[0075] O A =Attention(Q A A T +PE)Attention(AK AT +PE)V A +conv(V A ) (13)

[0076] Step 2.10: Add the agent attention value to the input value to get the residual calculation result And use the result as a multi-layer perceptron mlp AE The outputs of the calculations are added to get the encoder output of the proxy attention.

[0077]

[0078] Step 2.11: Use attention aggregation and multi-layer perceptron method to perform multi-task classification. The calculation method of attention aggregation is shown in formula (15), and the calculation result S L The last dimension is embedded in the encoding dimension for aggregation. The result is calculated by the classifier defined by the multi-layer perceptron for different tasks to obtain the classification result logit form 、logit domain 、logit peculiarity As shown in formula (16). And use the result as the input of the cross entropy loss function to enter the processing process of step 3. aggr Multilayer perceptron, mlp, is the result aggregation form , mlp domain , mlp peculiarity Multilayer perceptrons for classification of three tasks respectively.

[0079] S L =softmax(mlp aggr (O AE ))×O AE (15)

[0080]

[0081] Step 3 includes the following sub-steps:

[0082] Step 3.1: The model adopts a combined loss function that includes three sub-task loss functions. The loss function of the sub-task is defined by the cross entropy loss function, and the output of each loss function is defined as L domin 、L form and L unique The combined loss function is defined as:

[0083]

[0084] Where α, β and γ are the linear combination coefficients in the combined loss function.

[0085] Step 4 includes the following sub-steps:

[0086] Step 4.1: The user inputs the case description text, and the model outputs the predicted subject area, object form, and originality classification probability value p'(y fc )、p'(y dc ) and p'(y pc ). And define the parameter vectors λ and η, and perform linear operations with the prediction results to form the final prediction results.

[0087]

[0088] Step 4.2: Obtain the classification result based on the maximum classification value mapping obtained from each subtask.

[0089] Step 4.3: Calculate the entropy H(y) of the task by predicting the probability of each subtask as shown in formula (19), where p(y fc )、p(y dc ) and p(y pc ) are the prediction results of the three subtasks respectively.

[0090]

[0091] Step 4.4: After each prediction, calculate the partial derivatives of H(y) with respect to λ and η, update the values ​​of λ and η, and use them for the next prediction. As shown in formula (20), this achieves the goal of enhancing the model prediction confidence, reducing the entropy value, and improving the prediction accuracy during the model prediction process.

[0092]

[0093] While specific embodiments of the present invention have been disclosed for illustrative purposes, intended to facilitate understanding and implementation of the present invention, those skilled in the art will appreciate that various substitutions, variations, 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 disclosure of the preferred embodiments, and the scope of protection claimed in the present invention shall be determined by the scope of the claims.

Claims

1. A method for identifying the scope of protection of subject matter in a copyright case, comprising the following steps: 1) Obtain multiple copyright infringement cases and annotate them to obtain a copyright infringement case dataset; The annotation information includes the scope of the field to which the work belongs, the object form of the work, and the originality of the work; 2) establishing a copyright object protection scope determination model to predict the scope of the work, the object form, and the originality of the work; the copyright object protection scope determination model includes an encoder module, a joint attention calculation module, a splicer module, a multi-layer perceptron module, and a classifier module; 3) Using the copyright infringement case dataset to train the copyright object protection scope determination model; wherein, during the training process, the encoder module encodes the input work description text W and its annotation information to obtain a text vector, the text vector including the work description text vector W n , the object's domain category label vector L dv , object form category label vector L fv And the work originality category label vector L sv ; The joint attention calculation module calculates W v With L dv The multi-head attention vector between W v With L fv The multi-head attention vector between and W v With L sv The multi-head attention vector W between att|Ls ; The splicer module combines the three multi-head attention vectors W att|Ls Splice to form a vector W concate ; The joint attention calculation module converts the vector W concate Respectively with the query weight matrix W Q , key weight matrix W K , value weight matrix W V Multiply and calculate Q A =W Q W concate , K A =W K W concate 、V A =W V W concate ; The multi-layer perceptron module mlp A Q A The proxy token A is calculated as input; the encoder module uses the relative position encoding calculation method to encode the work description text W to obtain the encoding vector PE; the joint attention calculation module is based on the encoding vector PE, proxy token A, Q A , K A 、V A The agent attention value O is calculated A , and then the agent attention value O A With vector W concate Add up to get the residual calculation result and will The output result obtained by inputting the multilayer perceptron module is Add together to get O AE ; Then calculate the aggregation result S L =softmax(mlp aggr (O AE ))×O AE , the classifier module is based on the aggregation result S L Predict the work description text W to which the work belongs, and the prediction result logit domain , Work object form prediction result logit form And the prediction result logit of the originality of the work peculiarity ; Then, the loss value is calculated based on the prediction results and the annotation results to optimize the copyright object protection scope determination model; 4) Inputting the work to be identified into the trained copyright object protection scope determination model to obtain the domain scope, object form and originality of the work.

2. The method according to claim 1, characterized in that Using multi-layer perceptron module mlp A , with Q A The proxy token A is calculated for the input; among them, the multi-layer perceptron module Linear is a linear function and ReLU is an activation function.

3. The method according to claim 1 or 2, characterized in that Using the agent attention calculation formula O A =Attention(Q A A T +PE)Attention(AK AT +PE)V A +conv(V A ) Calculate the agent attention value O A , where conv is a one-dimensional convolutional neural network.

4. The method according to claim 1, wherein Use the cross entropy loss function to predict the result logit according to the scope of the work domain The loss value L is calculated based on the scope of the field to which the annotated work belongs. domin , using the cross entropy loss function to predict the result logit according to the object form of the work form The loss value L is calculated based on the object form of the annotated work form , using the cross entropy loss function to predict the result logit based on the originality of the work peculiarity The loss value L is calculated based on the originality of the marked work unique ; Then calculate the total loss value L = αL domin +βL form +γL unique , α+β+γ=1, and the copyright object protection scope determination model is optimized according to the total loss value L.

5. The method according to claim 1, wherein In step 4), define the parameter vectors λ and η; the predicted domain range p'(y fc ), object form p'(y dc ) and originality p'(y pc ) to perform linear operations Then calculate the prediction entropy Then calculate the partial derivatives of H(y) with respect to λ and η respectively The updated λ and η are used for the next prediction.

6. The method according to claim 1, 2 or 3, characterized in that: The copyright infringement cases mentioned include civil copyright infringement cases and criminal copyright infringement cases.

7. The method according to claim 1, characterized in that The encoder module uses a pre-trained language model as a text embedding encoder.

8. A server, characterized in that: The method 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 7.

9. 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 7 is implemented.

Images