Intellectual property infringement case matching method and system based on federal meta-learning

Abstract

The application discloses a kind of based on federal meta-learning's intellectual property infringement case matching method and system, belong to natural language processing technical field.The application is by constructing the five-level architecture and corresponding function module of data layer, federal learning layer, semantic enhancement layer, meta-learning adaptation layer and application layer composition, to realize infringement case case search.Improvements include: optimizing the parameter aggregation efficiency of federal learning, improving the small sample adaptation speed of meta-learning, innovating the feature extraction mode of legal semantics, adding the result explainability generation module, to solve the problems of data privacy cooperation, cross-scene adaptation, legal semantic understanding, search efficiency and result explainability, etc.

Description

Technical Field

[0001] This system belongs to the field of natural language processing technology, specifically a... base The intellectual property infringement case matching method and system learned by Federal Yuan. Background Technology

[0002] In judicial practice of intellectual property infringement cases, case retrieval is a core step in achieving consistent judgments in similar cases and improving adjudication efficiency. For judges, it is necessary to refer to the infringement determination standards, legal application logic, and compensation scales of historical similar cases to avoid judicial deviations; for rights holders, it can clarify the direction of rights protection and the focus of evidence, reducing the cost of rights protection; for the judicial system, it can reduce redundant research, optimize resource allocation, and shorten the case handling cycle.

[0003] However, current intellectual property infringement case retrieval faces the following four problems: 1. Data security and privacy issues: The information items for case retrieval mainly include court documents, rights protection materials and appraisal reports. However, these data are scattered among courts, rights holders and appraisal institutions. Because the materials involve the privacy or trade secrets of the parties involved and cannot be disclosed, the training samples for the model are insufficient and the model accuracy is low.

[0004] 2. Weak adaptability across different types of cases: Infringement cases involve a wide variety of types and each type of case has a different focus. Traditional models need to be trained separately for each type of infringement case, which is time-consuming and laborious. In addition, some case types have too few samples to train the model.

[0005] 3. Existing models have poor understanding of legal semantics: Judgments in tort cases contain specific professional terms, and general models cannot accurately link the facts of infringement with legal requirements, resulting in large semantic matching errors.

[0006] 4. Low search efficiency: Manual search of individual cases can only be conducted based on simple keywords. The cases are not targeted and require a lot of time to manually judge whether the cases are referential.

[0007] In summary, how to protect textual privacy while breaking through data barriers, improving the accuracy of cross-type case matching and legal understanding while also taking efficiency into account are urgent issues to be addressed in tort case matching. Summary of the Invention

[0008] This invention discloses a base The intellectual property infringement case matching method and system developed by Federal Yuan can break down data collaboration barriers, improve matching accuracy in niche scenarios, ensure that the recommended results conform to the judgment logic, and meet the high timeliness requirements of judicial and rights protection.

[0009] To achieve the above objectives, the technical solution of the present invention includes the following:

[0010] A sort of base The intellectual property infringement case matching method in the Federal Meta-learning framework applies an institutional node, and the method includes: Preprocess the text data of the infringement cases to be matched; The preprocessed text data and the text data of historical similar cases are input into the case matching model to obtain the case matching result of the infringement case to be matched. The case matching model includes: A general feature extractor is used to compute the legal semantic vector of preprocessed text data; The scene-aware fusion layer is used to generate scene features of the infringement case to be matched based on the feature words in the preprocessed text data, and to fuse the legal semantic vector and scene features. The task output layer is used to obtain the case matching result of the infringement case to be matched based on the similarity between the fused feature vector of the infringement case to be matched and the fused feature vector of historical similar cases.

[0011] Furthermore, the preprocessed text data for infringement cases includes: The SimHash algorithm is used to deduplicate the original text data of infringement cases, and the deduplicated text data is then logically segmented. Through machine pre-annotation and manual verification, key information in each logical segment is prioritized; the key information includes: infringement type and legal provisions. The text data of each logical segment is anonymized, and semantic integrity verification is used to ensure that the anonymized text data is semantically consistent with the original text data.

[0012] Furthermore, the process of training the case matching model includes: Construct a meta-task set; where each meta-task Supports collection The query set contains several labeled samples within a group. Includes other samples in this group; An initial case matching model is generated, wherein the general feature extractor in the initial case matching model is obtained by federated learning training of a pre-trained model; The case matching model is iteratively trained using an inner and outer loop based on a meta-task set; wherein, the inner loop training refers to freezing the parameters of the general feature extractor and training based on the support set. Adjusting the parameters of the scene-aware fusion layer and the task output layer; the outer loop training refers to training based on the query set. Adjust the parameters of the general feature extractor, the scene-aware fusion layer, and the task output layer; The node types in the infringement case scenario knowledge graph include: infringement type, legal elements, and feature words. The edge types in the infringement case scenario knowledge graph include: belonging, containing, requiring proof, corresponding legal provisions, and core features. The content of the infringement type includes: trademark infringement, patent infringement, copyright infringement, trade secret infringement, and luxury goods exclusive infringement. The content of the legal elements includes: likelihood of confusion, substantial similarity, likelihood of access, subjective fault, and damage result. The content of the feature words includes: trademark logo, source code, design drawings, product appearance, brand logo, and customized process.

[0013] Furthermore, federated learning training is performed on the pre-trained model, including: According to the global model parameters issued by the federal server Generate local model ;in, Indicates the federal learning and training rounds. Indicates the index of the organization node; Training a local model based on local data ; Gradient of model parameters obtained during training Uploaded to the federated server, so that the federated server can aggregate the model parameter gradients uploaded by each agency node. To update the global model and set the updated global model parameters After being distributed to the various institutional nodes, the command And re-execute the global model parameters issued by the federal server. Generate local model ; The process continues until the preset conditions are met, resulting in a pre-trained model.

[0014] Furthermore, train the fusion feature vector generation model. loss function ;in, Represents the first weight coefficient, cosine similarity loss. , The semantic vector of the target text data. The semantic vectors representing candidate text data, and the triple alignment loss. , A semantic vector representing an infringement. A semantic vector representing legal requirements. The semantic vector representing the cited legal provision. Indicates the temperature coefficient. , , These are the first weight, the second weight, and the third weight, respectively.

[0015] Furthermore, the gradient of the model parameters The generation process includes: Obtaining the fusion feature vector generation model The gradient of the original parameters at time; Semantic parsing and mapping of the original parameter gradients are performed to identify the semantic information corresponding to each gradient dimension; Different importance weights are assigned to different gradient dimensions based on the importance of this semantic information; Based on the importance weights, noise of appropriate intensity is added to the corresponding gradient dimensions to obtain the model parameter gradients. .

[0016] Furthermore, the federated server aggregates the model parameter gradients uploaded by each agency node. To update the global model, including: Calculate the weights of the mechanism nodes Wherein, the weight ,in, For institutional nodes The effective sample size For all institutional nodes The number of valid samples, Indicates institutional nodes Text quality score, scene coverage score Legal semantic consistency factor The tripartite combination includes: the tortious act, the legal elements, and the cited legal provisions. For weight adjustment parameters, For institutional nodes The fusion feature vectors are used to generate model parameters; Based on weight gradient of model parameters Perform a weighted average to obtain the update amount of the global model. ; Based on the update amount Update the global model.

[0017] Furthermore, when dealing with a new infringement scenario, the case matching model is adjusted through the following steps: Several annotated court judgment samples were collected from this new infringement scenario to form a support set. ; Based on support set The case matching model is trained in an inner loop to obtain the adjusted case matching model.

[0018] Furthermore, based on the legal semantic vector and scene features of the text data, a fused feature vector is obtained, including: By calculating the correlation coefficient between the legal semantic vector and the scene feature vector, the legal semantic weight and the scene feature weight are obtained. Based on the legal semantic weight and the scene feature weight, the legal semantic vector and the scene feature are weighted and calculated to obtain the fused feature vector.

[0019] A sort of base The intellectual property infringement case matching system for Federal Meta-Learning includes: The preprocessing module is used to preprocess the text data of the infringement cases to be matched; The case matching model is used to obtain the case matching result of the infringement case to be matched based on the preprocessed text data and the text data of historical cases. The case matching model includes: A general feature extractor is used to compute the legal semantic vector of preprocessed text data; The scene-aware fusion layer is used to generate scene features of the infringement case to be matched based on the feature words in the preprocessed text data, and to fuse the legal semantic vector and scene features. The task output layer is used to obtain the case matching result of the infringement case to be matched based on the similarity between the fused feature vector of the infringement case to be matched and the fused feature vector of historical similar cases.

[0020] Compared with the prior art, the present invention has at least the following beneficial effects.

[0021] 1. Address data security and privacy issues to achieve cross-organizational compliance collaboration.

[0022] Existing technologies cannot share infringing data across institutions due to the presence of privacy and trade secrets, resulting in limited model training samples and low accuracy. This invention achieves data collaboration among institutions under compliance by encrypting parameter transmission and retaining original data locally. The model training samples cover all infringement scenarios, and the matching accuracy is improved to over 85%. At the same time, it avoids the leakage of sensitive data and ensures the legality and security of cross-institutional collaboration.

[0023] 2. Solve the problem of adapting to small samples in niche scenarios, and cover the retrieval of all infringement scenarios.

[0024] Existing technologies require separate model training for each infringement scenario, resulting in low accuracy for niche scenarios due to insufficient samples. This invention utilizes meta-learning and small-sample fine-tuning, requiring only 30 or fewer labeled texts to complete adaptation for niche scenarios. It achieves cross-scenario matching accuracy exceeding 85%, covering as many infringement scenarios as possible to meet the case retrieval needs of various types of infringement cases and reduce judicial bias caused by incomplete scenario coverage.

[0025] 3. Improve the accuracy of legal semantic understanding and avoid being misled by superficial textual similarities.

[0026] Existing technologies, based on general text processing, cannot accurately identify infringement-specific legal semantics such as the possibility of confusion and substantial similarity, resulting in semantic matching errors exceeding 30%. This invention, through fine-tuning of a legal pre-trained model and a scene attention mechanism, establishes a precise association between descriptions of infringement, legal elements, and cited legal provisions, reducing semantic matching errors to below 10%. The recommended similar cases align with the core needs of adjudication, enhancing the consistency and fairness of judicial decisions.

[0027] 4. Significantly improves search efficiency, adapting to the timeliness requirements of judicial and rights protection.

[0028] Existing technologies require lengthy manual searches for individual cases, and traditional models can take several hours to retrieve 100 texts in batches. This invention, through lightweight modeling and structured output, reduces the time required for single-case matching to within seconds and the time required for batch retrieval of hundreds of infringing texts to less than 10 minutes, meeting the need for rapid rights protection, optimizing the allocation of judicial resources, and reducing the time cost of rights protection. Attached Figure Description

[0029] Figure 1 base Architecture diagram of the intellectual property infringement case matching system learned by Federal Yuan. Detailed Implementation

[0030] The system will now be described in further detail with reference to the accompanying drawings. The examples given are for illustrative purposes only and are not intended to limit the scope of the system.

[0031] This invention addresses the need for case matching in infringement cases by making multi-dimensional core improvements to existing technologies, integrating three core technical architectures: legal semantic enhancement, federated learning, and meta-learning. Figure 1 As shown, it achieves infringement case retrieval through a five-level architecture consisting of a data layer, a federated learning layer, a semantic enhancement layer, a meta-learning adaptation layer, and an application layer, along with corresponding functional modules. Key improvements include: optimizing the parameter aggregation efficiency of federated learning, improving the small-sample adaptation speed of meta-learning, innovating the feature extraction method of legal semantics, and adding a result interpretability generation module, thereby specifically addressing issues such as data privacy collaboration, cross-scenario adaptation, legal semantic understanding, retrieval efficiency, and result interpretability.

[0032] 1. Data layer: Preprocessing and de-identification of multi-source data.

[0033] Each participating organization uploads the text data of infringement cases locally and performs deduplication, noise reduction, anonymization, segmentation, and annotation through a dedicated preprocessing workflow, ultimately generating standardized and anonymized preprocessed text. All preprocessed text is stored locally, and only the extracted relevant feature information is used for subsequent processing, laying the foundation for semantic extraction and feature fusion.

[0034] (1) Data preprocessing: Based on the traditional process, a feature priority labeling module is added. The SimHash algorithm is used to remove duplicates (threshold 0.95) and filter noise, and the data is segmented according to logic such as "investigation and findings" and "points of contention". Finally, through machine pre-labeling and manual verification, priority is assigned to key information such as infringement type and legal provisions (P1 core, P2 secondary, P3 auxiliary) to strengthen the model's focus on core features.

[0035] (2) Text Desensitization: Innovatively employing scenario-based dynamic desensitization technology. First, sensitive information types are identified based on the infringement scenario map, and then differentiated desensitization intensities are set: in judicial scenarios, the parties' surnames and the last four digits of the case number are retained; in rights protection scenarios, the description of the infringing act is focused; and in appraisal scenarios, information about the institution and personnel is masked. Simultaneously, a semantic integrity verification module ensures that the core semantics are not lost after desensitization (completeness). 98%.

[0036] 2. Federated Learning Layer: Federated Collaborative Training with Enhanced Legal Semantics.

[0037] Based on the preprocessed feature data, each node trains its local sub-model using a lightweight training framework based on its local fusion vector, and optimizes the model using an infringement triplet alignment loss function. After training, only the gradients of the model parameters are extracted, protected by gradient semantic differential privacy technology, and then uploaded to the federated server. The federated server aggregates the parameter gradients of all nodes according to a four-factor dynamic weight strategy, generates global model parameters, and distributes them to each node. This process is iterated until the global model accuracy reaches 85%.

[0038] (1) Infringement triple alignment loss function.

[0039] To address the shortcomings of traditional loss functions that only focus on surface text similarity, this invention introduces a legal logic weighting factor for improvement. Based on the traditional text similarity loss function, an infringement triplet alignment loss is added to construct a joint loss function, enabling the model to simultaneously consider both surface text similarity and deep legal logic connections. This improvement increases the model's focus on legal elements by 40%, effectively reducing misjudgments caused by surface text similarity.

[0040] Its core logic is to define the infringing act, legal elements, and cited legal provisions as a tort triple, and to design a loss function around the semantic consistency among these three elements. By constructing semantic vectors for each element in the triple and using vector distance to quantify their semantic relevance, the joint loss function formula is as follows: in For cosine similarity loss, Set to 0.8, This represents the triple alignment loss.

[0041] The calculation formula is: in For the semantic vector of the target case, The semantic vector of the candidate case.

[0042] The calculation formula is: in, The distance to the boundary is set to 0.2. , , The weights for the three components are set to 0.4, 0.4, and 0.2 respectively. For each element in the triple, construct a semantic vector for the infringement, legal requirements, and cited legal provisions.

[0043] (2) Multi-factor dynamic weight federated aggregation.

[0044] To address the issue of fixed weight allocation in traditional federated aggregation, this invention introduces an improved real-time contribution evaluation mechanism: a four-factor dynamic weight aggregation mechanism. This mechanism comprehensively evaluates the data contribution of each node and monitors changes in data quality in real time. When a node's text quality score... If the weight is below 0.6 for three consecutive rounds, the system will automatically reduce its weight ratio to avoid interference from poor-quality data on the global model. This improvement increases the convergence speed of the global model by 30% and demonstrates higher stability in multi-agency collaborative scenarios.

[0045] Specifically, first, after each node completes local training, it updates the model parameters (gradients). After gradient semantic differential privacy protection, the data is uploaded to the server.

[0046] Next, the server calculates the aggregate weight for each node: The parameters in the above formula are explained as follows: Sample size factor : For nodes The effective sample size To determine the number of valid samples for all nodes, the validity of each node's samples must first be verified, and duplicate cases and cases with incomplete information (such as cases lacking descriptions of infringement) must be removed. The scale contribution of node data must also be quantified. Text quality score The system automatically scores the input text based on preset rules, considering three aspects: whether the text contains core information, whether the expression conforms to the format of judicial documents, and whether the legal terminology is used correctly. The score range is between 0 and 1. Scene coverage score : ; Legal semantic consistency factor : The higher the logical fit of the infringing triples in the node data, the better. The closer to 1, This is to assess the contribution of node data to legal reasoning; Weighting parameters The initial parameters are all set to 0.25, and can be adjusted according to your needs. For nodes Local model parameters.

[0047] Then, the server uses the calculated weights The gradients uploaded by all nodes are weighted and averaged to obtain the update amount of the global model. = .

[0048] Finally, the server uses Update the global model parameters, and then distribute the updated global model to each participating node.

[0049] (3) Gradient semantic differential privacy protection.

[0050] To further enhance data privacy protection during gradient transmission in federated learning, this invention introduces a gradient semantic differential privacy technique. This technique abandons the traditional differential privacy method of applying uniform noise to all gradient dimensions, and instead constructs a differentiated noise injection process based on semantic importance.

[0051] Specifically, this technique first performs semantic parsing and mapping on the original gradients to identify the semantic information corresponding to each gradient dimension; then, it assigns differentiated importance weights to different gradient dimensions based on their semantic importance; finally, it adds noise of appropriate intensity according to the assigned weights. The gradients processed in this way are then transmitted through an encrypted channel.

[0052] This method implements precise semantic-level noise perturbations at the gradient level, while strictly meeting judicial data privacy protection standards (privacy budget). Under the premise of minimizing noise interference with core legal semantic features, the use of the Laplace noise mechanism ensures the rationality of noise addition. This effectively balances the inherent contradiction between the strength of privacy protection and model performance, achieving a unity between privacy security and model utility.

[0053] 3. Semantic Enhancement Layer: Legal semantic extraction and scene feature fusion.

[0054] In this stage, the semantic enhancement layer extracts legal semantics and fuses scene features into the preprocessed text. First, a fine-tuned pre-trained legal model is used to generate legal semantic vectors, and a judicial keyword attention mechanism is introduced to strengthen the expression of core semantics. Then, scene features are extracted based on the infringement scene graph and transformed into feature vectors. Finally, the two types of vectors are concatenated according to preset weights to form a fusion vector with a unified dimension.

[0055] (1) Construction of infringement scenario map.

[0056] This invention constructs a legal knowledge graph comprising three core node categories: infringement type, legal elements, and characteristic words, along with their associated relationships. The construction process of this graph includes three core steps: knowledge extraction, relationship definition, and graph optimization. The specific process is as follows: 1) Knowledge extraction: Based on laws and regulations such as the Trademark Law and the Patent Law, as well as a large number of judicial documents, the Named Entity Recognition technology is used to automatically extract three types of nodes: infringement type, legal elements and feature words, with an extraction accuracy rate of over 92%. 2) Relationship definition: Based on legal logic and judicial practice, five core semantic relationships are clearly defined: “belonging to”, “including”, “requiring proof”, “corresponding legal provisions”, and “core features”, in order to accurately depict the inherent connections between nodes; 3) Graph optimization: A combination of manual verification and machine learning error correction is used to correct erroneous associations generated during the automatic extraction process, thereby ensuring the accuracy and reliability of graph knowledge.

[0057] The core content of the constructed infringement scenario map is shown in Table 1: Types of infringement Trademark infringement, patent infringement, copyright infringement, trade secret infringement, and infringement of exclusive rights to luxury goods, etc. Legal requirements Possibility of confusion, substantial similarity, likelihood of contact, subjective fault, and resulting damage. Feature words Trademarks, source code, design drawings, product appearance, brand logo, custom processes, etc. Table 1 (2) Fine-tuning of the legal pre-training model.

[0058] To address the issue of insufficient adaptation of general legal models to infringement scenarios, this invention improves them through scenario-based incremental pre-training. Based on locally labeled data, a secondary pre-training of basic legal pre-trained models (such as BERT-LAW and LawBERT) is performed. By introducing a corpus specifically for infringement scenarios (containing 100,000 infringement judgments) and employing a strategy combining global learning and local fine-tuning, 60% of the model's bottom-level parameters are fixed, while training is performed only on the top-level 40% of the parameters and the newly added scenario adaptation layer. This approach improves scenario adaptation accuracy while avoiding semantic shifts.

[0059] Data labeling: Labeling text data and features such as infringement facts and legal basis.

[0060] Local model fine-tuning: The local fine-tuning mode is adopted, which fixes the low-level parameters of the model and trains only the high-level parameters of the model to prevent semantic shift caused by overtraining.

[0061] Introducing a judicial keyword attention mechanism: A judicial keyword attention module is added to the model's encoding layer, assigning higher weights to keywords related to the judiciary and torts. The calculation formula is as follows: in This represents the hidden state of the j-th token. is a learnable parameter for the attention mechanism, and n is the total number of tokens in the text. The larger the value, the more important the token's contribution to legal semantic extraction.

[0062] Performance verification: using the validation set Evaluate the accuracy of the fine-tuned model. For ease of evaluation, classification accuracy (acc) is primarily used as the evaluation metric, calculated using the following formula: Set the verification threshold to 85%, if If the requirements are not met, adjust the model parameters and retrain until the model is satisfactory.

[0063] (3) Scene feature encoding and fusion.

[0064] This invention proposes a dynamic weighted fusion improvement mechanism to construct a system for the collaborative fusion of scene features and legal semantics. Unlike traditional fixed-weight fusion methods, this mechanism dynamically adjusts the fusion weights by calculating the correlation coefficient between the legal semantic vector and the scene feature vector in real time: when the correlation is high (≥0.8), the legal semantic weight is set to 0.7 and the scene feature weight is set to 0.3; when the correlation is low (<0.5), both weights are set to 0.5. This ensures that the fused vector accurately reflects the core features under different infringement scenarios. The specific operation steps are as follows: Constructing a scene-specific vocabulary: This involves identifying various scenarios in infringement cases, collecting typical characteristic words for each scenario, and constructing a scene-specific vocabulary. ; Feature vector transformation: Converting scene-specific words in the text into feature vectors of appropriate dimensions. To ensure that it has sufficient semantic distinguishing ability; Feature fusion: combining scene feature vectors and legal semantic vector The fused vector is formed by weighted concatenation according to the following formula. in and These are the weights of the legal semantic vector and the scene feature vector, respectively. ; Noise filtering: Based on keyword density and semantic relevance, sentences in the text are filtered to remove redundant expressions, effectively reducing the interference of noise on the quality of feature vectors.

[0065] 4. Meta-learning adaptation layer: cross-scene adaptation based on infringement scene graph.

[0066] The meta-learning adaptation layer loads specialized model parameters optimized by the semantic enhancement layer and constructs a meta-task set based on the infringement scenario graph, classifying infringement cases into different scenario types. The outer loop optimizes the model's initial parameters, enhancing its learning ability in general scenarios; for specific niche scenarios, the inner loop mechanism uses a small support set (only 30-50 labeled texts) to quickly fine-tune the model, achieving efficient and accurate scenario adaptation.

[0067] 1) Adaptive model hierarchical network structure.

[0068] To achieve fast and accurate adaptation in scenarios with small sample sizes, this invention designs an adaptation model that includes three functional levels.

[0069] The model's bottom layer is a general feature extractor that directly reuses the global legal semantic model trained through federated learning. Most parameters in this layer are fixed or allowed only minimal fine-tuning in subsequent processes. This aims to encode the general legal language rules and basic logic learned by the model from massive amounts of data into a stable knowledge base, thereby ensuring that the model does not lose its fundamental semantic understanding ability when quickly adapting to new scenarios and effectively avoiding catastrophic forgetting.

[0070] The middle layer of the model is the scene-aware fusion layer, which is the core innovative module of this invention. It introduces a cross-attention mechanism to deeply fuse feature vectors representing specific scenarios (such as luxury goods infringement) extracted from the infringement scenario knowledge graph with the general legal semantic features output from the lower layer. This mechanism uses semantic features as queries and scene features as keys and values, dynamically calculating association weights. This allows the model to adaptively focus on the legal elements most relevant to the current scenario (e.g., focusing on "confusion probability" in trademark infringement), thereby generating fused features rich in scenario information.

[0071] The top layer of the model is a lightweight task output layer, serving as a flexible task interface. During the meta-training phase, to comprehensively improve the model's generalization ability, this layer structure is configured according to different proxy tasks: for example, a Softmax classifier is used in scene classification tasks to drive the model to learn the essential features that distinguish different scenes; a similarity calculation module is used in class matching tasks. In the final application adaptation phase for new scenarios, this layer is uniformly fixed as the similarity calculation module (such as a cosine similarity layer) required for class matching, so that the parameters of the middle and top layers can be quickly fine-tuned with a very small number of samples, enabling the model to obtain high-precision matching capabilities for the specific scene.

[0072] 2) Construction of the meta-task set.

[0073] Historical infringement cases are grouped based on their semantic similarity to construct a set of training tasks (metatatata task set) for meta-learning. Each metatatata task is defined as follows: ,in It is the support set, which contains a small number of labeled samples in this group, used to simulate and perform rapid adaptation of the model on new tasks; The query set, containing other samples within the same group, is used to evaluate the generalization performance of the adapted model and provide feedback for the meta-optimization of the outer loop. During scene grouping, cosine similarity is used to measure the feature vectors of cases, and the similarity is... Scenarios are grouped together to ensure that tasks within the group have sufficient semantic relevance, thereby improving the efficiency of meta-knowledge transfer.

[0074] 3) Dual-loop meta-training mechanism.

[0075] To train the above model, this invention employs a dual-loop meta-training mechanism, the goal of which is to enable the model to learn how to learn quickly.

[0076] Inner loop (rapid adaptation within the task): This stage simulates the rapid adaptation process of the model when facing new scenarios. For a meta-task The model starts with a set of initial parameters and performs only a few gradient updates on the parameters of the middle scene fusion layer and the top task layer. The update objective is to minimize the gradients on the task support set. This process is highly efficient and requires only 30-50 samples to achieve effective adaptation.

[0077] Outer Loop (Meta-optimization and Generalization): This stage aims to optimize the initial parameters of the model, giving it good potential for rapid adaptation. Specifically, the system collects all meta-tasks after adaptation in the inner loop, and then applies them to their respective query sets. The model's performance (loss) is evaluated, and all parameters (including underlying basic parameters) are updated using gradient descent. This allows the initial state obtained by the model to achieve excellent performance with minimal inner loop fine-tuning when facing new future scenarios.

[0078] Iterative process: The above-mentioned inner and outer loop processes are repeated on the constructed large meta-task set until the model obtains a stable and rapid adaptive capability.

[0079] 4) New scenario adaptation application process.

[0080] When the system needs to handle a completely new, niche infringement scenario (such as "luxury goods design plagiarism"), the complete adaptation process is as follows: Construct a small support set. Collect 30-50 labeled court judgment samples from this new scenario to form a support set.

[0081] Initiate rapid fine-tuning within the inner loop. Load a model with strong adaptive potential after meta-training, using it as a starting point. Fix the bottom-level general feature extractor, and only fine-tune the parameters of the mid-level scene fusion layer and the top-level similarity calculation layer using a small sample support set of the new scene. This process typically converges quickly, completing in about 10 iterations.

[0082] Validation and Deployment. Evaluate the matching accuracy of the adapted model on the reserved validation set for this scenario. When the accuracy reaches the preset standard, the adaptation of the new scenario is considered complete, and the model can be deployed in actual case retrieval applications.

[0083] 5. Application layer: Case matching and interpretable result output.

[0084] Users input the text of a tort case to be matched into the judicial case handling system. The system then transmits this text to the semantic enhancement layer and the adapted model. The application layer calls the adapted model to calculate the similarity between the text to be matched and the fused vectors of historical similar cases, and performs secondary optimization and ranking based on the legal semantic matching degree. Simultaneously, the system extracts structured judgment points from historical similar cases using the TextRank algorithm and generates interpretable natural language matching reasons. Finally, the results are output in the form of numerical scores, textual reasons, and structured information.

[0085] (1) Similarity calculation and ranking.

[0086] Input the text of the infringement case to be matched and convert it into a fusion vector. By calculating the fusion vector between this vector and the historical case text, The cosine similarity between them is used to obtain the matching score. The calculation formula is as follows: in The value ranges from [0,1], with higher values ​​indicating a higher degree of case matching.

[0087] (2) Case Ranking: To improve the comprehensiveness of the case recommendation results, a second optimization ranking is performed based on the basic similarity ranking, using legal semantic matching degree. Legal semantic matching degree is defined as the semantic similarity between the text to be matched and historical cases at the level of legal elements. The system prioritizes recommending related cases with a legal element semantic similarity of not less than 70% to ensure that the recommendation results are more in line with the judgment logic.

[0088] (3) Extracting key points of judgment: Based on the TextRank algorithm, and combining the semantic importance and contextual relevance of sentences, the core key points of judgment are automatically extracted from historical similar case texts. The extracted content is structured and organized into four categories: basis for infringement determination, cited legal provisions, compensation standards, and judgment rules, forming a clear and organized judgment summary, avoiding the simple piling up of the original text.

[0089] (4) Results output: The case matching function supports highlighting keywords that are highly related to the text to be matched in the historical case text. All matching results (including similarity scores, structured judgment points and matching reasons) can be exported as standardized format documents, which facilitates subsequent analysis and archiving, and enables rapid response and system integration.

[0090] In summary, the data layer of this invention integrates multi-source data and forms standardized samples through deduplication, noise reduction, structuring, and contextual dynamic desensitization, thereby ensuring privacy and security.

[0091] The federated learning layer of this invention adopts a legal semantic enhancement federated learning framework, ensures learning quality through the infringement triple alignment loss function, optimizes collaborative efficiency by using four-factor dynamic weight aggregation, and introduces gradient semantic differential privacy technology to ensure data transmission security. Finally, a general basic legal semantic model is trained, thereby solving the problem of data security and collaboration contradiction in the prior art.

[0092] The semantic enhancement layer of this invention starts with the basic model produced by federated training, performs scenario-based incremental pre-training and local fine-tuning on the infringement scenario corpus, and introduces a judicial keyword attention mechanism and dynamic weight fusion strategy to construct an accurate legal semantic-scenario feature fusion vector, thereby solving the problem of legal semantic understanding bias in the prior art.

[0093] The meta-learning adaptation layer of this invention is based on a specialized model produced by the semantic enhancement layer. It uses the infringement scene graph to construct meta-tasks and, through a dual-loop training mechanism, endows the model with the meta-ability to quickly adapt to new scenes with a small number of samples, thereby solving the problem of insufficient cross-scene semantic adaptation in the prior art.

[0094] The application layer of this invention calculates similarity and performs secondary ranking in conjunction with legal semantics. It extracts structured judgment points through the TextRank algorithm and generates natural language matching reasons, thereby solving the problems of disconnect between retrieval efficiency and timeliness, and poor interpretability of results.

[0095] Although specific embodiments of the system have been disclosed for illustrative purposes to aid in understanding and implementing the system, those skilled in the art will understand that various substitutions, variations, and modifications are possible without departing from the spirit and scope of the system and the appended claims. Therefore, the system should not be limited to the content disclosed in the preferred embodiments, and the scope of protection claimed by the system is determined by the scope defined in the claims.

Claims

1. A method for matching intellectual property infringement cases based on federated meta-learning, characterized in that, The method, which applies an institutional node, includes: Preprocess the text data of the infringement cases to be matched; The preprocessed text data and the text data of historical similar cases are input into the case matching model to obtain the case matching result of the infringement case to be matched. The case matching model includes: A general feature extractor is used to compute the legal semantic vector of preprocessed text data; The scene-aware fusion layer is used to generate scene features of the infringement case to be matched based on the feature words in the preprocessed text data, and to fuse the legal semantic vector and scene features. The task output layer is used to obtain the case matching result of the infringement case to be matched based on the similarity between the fused feature vector of the infringement case to be matched and the fused feature vector of historical similar cases. The process of training the case matching model includes: A meta-task set is constructed based on a knowledge graph of infringement case scenarios; where each meta-task... Supports collection The query set contains several labeled samples within a group. Includes other samples in this group; An initial case matching model is generated, wherein the general feature extractor in the initial case matching model is obtained by federated learning training of a pre-trained model; The case matching model is iteratively trained using an inner and outer loop based on a meta-task set; wherein, the inner loop training refers to freezing the parameters of the general feature extractor and training based on the support set. Adjusting the parameters of the scene-aware fusion layer and the task output layer; the outer loop training refers to training based on the query set. Adjust the parameters of the general feature extractor, the scene-aware fusion layer, and the task output layer; The node types in the infringement case scenario knowledge graph include: infringement type, legal elements, and feature words; the edge types in the infringement case scenario knowledge graph include: belonging, containing, requiring proof, corresponding legal provisions, and core features; the content of the infringement type includes: trademark infringement, patent infringement, copyright infringement, trade secret infringement, and luxury goods exclusive infringement; the content of the legal elements includes: likelihood of confusion, substantial similarity, likelihood of access, subjective fault, and damage result; and the content of the feature words includes: trademark logo, source code, design drawings, product appearance, brand logo, and customized process. This includes federated learning training of the pre-trained model, which includes: According to the global model parameters issued by the federal server Generate local model ;in, Indicates the federal learning and training rounds. Indicates the index of the organization node; Training a local model based on local data ; Gradient of model parameters obtained during training Uploaded to the federated server, so that the federated server can aggregate the model parameter gradients uploaded by each agency node. To update the global model and set the updated global model parameters After being distributed to the various institutional nodes, the command And re-execute the global model parameters issued by the federal server. Generate local model ; The process continues until the preset conditions are met, resulting in a pre-trained model.

2. The method according to claim 1, characterized in that, The preprocessed text data of the infringement cases to be matched includes: The SimHash algorithm is used to deduplicate the original text data of infringement cases, and the deduplicated text data is then logically segmented. Through machine pre-annotation and manual verification, key information in each logical segment is prioritized; the key information includes: infringement type and legal provisions. The text data of each logical segment is anonymized, and semantic integrity verification is used to ensure that the anonymized text data is semantically consistent with the original text data.

3. The method according to claim 1, characterized in that, Training the fusion feature vector generation model loss function ;in, Represents the first weight coefficient, cosine similarity loss. , The semantic vector of the target text data. The semantic vectors representing candidate text data, and the triple alignment loss. , A semantic vector representing an infringement. A semantic vector representing legal requirements. The semantic vector representing the cited legal provision. Indicates the temperature coefficient. , , These are the first weight, the second weight, and the third weight, respectively.

4. The method according to claim 1, characterized in that, The gradient of the model parameters The generation process includes: Obtain the training fusion feature vector generation model The gradient of the original parameters at time; Semantic parsing and mapping of the original parameter gradients are performed to identify the semantic information corresponding to each gradient dimension; Different importance weights are assigned to different gradient dimensions based on the importance of this semantic information; Based on the importance weights, noise of appropriate intensity is added to the corresponding gradient dimension to obtain the model parameter gradient. .

5. The method according to claim 1, characterized in that, The federated server aggregates the model parameter gradients uploaded by various agency nodes. To update the global model, including: Calculate the weights of the mechanism nodes Wherein, the weight ,in, For institutional nodes The effective sample size For all institutional nodes The number of valid samples, Indicates institutional nodes Text quality score, scene coverage score Legal semantic consistency factor The tripartite combination includes: the tortious act, the legal elements, and the cited legal provisions. For weight adjustment parameters, For institutional nodes The fusion feature vectors are used to generate model parameters; Based on weight gradient of model parameters Perform a weighted average to obtain the update amount of the global model. ; Based on the update amount Update the global model.

6. The method according to claim 1, characterized in that, When dealing with a new infringement scenario, adjust the case matching model by following these steps: Several annotated court judgment samples were collected from this new infringement scenario to form a support set. ; Based on support set The case matching model is trained in an inner loop to obtain the adjusted case matching model.

7. The method according to claim 1, characterized in that, Based on the legal semantic vector and scene features of the text data, a fused feature vector is obtained, including: By calculating the correlation coefficient between the legal semantic vector and the scene feature vector, the legal semantic weight and the scene feature weight are obtained. Based on the legal semantic weight and the scene feature weight, the legal semantic vector and the scene feature are weighted and calculated to obtain the fused feature vector.

8. A kind base The intellectual property infringement case matching system developed by Federal Yuan Learning is characterized by, The system includes: The preprocessing module is used to preprocess the text data of the infringement cases to be matched; The case matching model is used to obtain the case matching result of the infringement case to be matched based on the preprocessed text data and the text data of historical cases. The case matching model includes: A general feature extractor is used to compute the legal semantic vector of preprocessed text data; The scene-aware fusion layer is used to generate scene features of the infringement case to be matched based on the feature words in the preprocessed text data, and to fuse the legal semantic vector and scene features. The task output layer is used to obtain the case matching result of the infringement case to be matched based on the similarity between the fused feature vector of the infringement case to be matched and the fused feature vector of historical similar cases. The process of training the case matching model includes: A meta-task set is constructed based on a knowledge graph of infringement case scenarios; where each meta-task... Supports collection The query set contains several labeled samples within a group. Includes other samples in this group; An initial case matching model is generated, wherein the general feature extractor in the initial case matching model is obtained by federated learning training of a pre-trained model; The case matching model is iteratively trained using an inner and outer loop based on a meta-task set; wherein, the inner loop training refers to freezing the parameters of the general feature extractor and training based on the support set. Adjusting the parameters of the scene-aware fusion layer and the task output layer; the outer loop training refers to training based on the query set. Adjust the parameters of the general feature extractor, the scene-aware fusion layer, and the task output layer; The node types in the infringement case scenario knowledge graph include: infringement type, legal elements, and feature words; the edge types in the infringement case scenario knowledge graph include: belonging, containing, requiring proof, corresponding legal provisions, and core features; the content of the infringement type includes: trademark infringement, patent infringement, copyright infringement, trade secret infringement, and luxury goods exclusive infringement; the content of the legal elements includes: likelihood of confusion, substantial similarity, likelihood of access, subjective fault, and damage result; and the content of the feature words includes: trademark logo, source code, design drawings, product appearance, brand logo, and customized process. This includes federated learning training of the pre-trained model, which includes: According to the global model parameters issued by the federal server Generate local model ;in, Indicates the federal learning and training rounds. Indicates the index of the organization node; Training a local model based on local data ; Gradient of model parameters obtained during training Uploaded to the federated server, so that the federated server can aggregate the model parameter gradients uploaded by each agency node. To update the global model and set the updated global model parameters After being distributed to the various institutional nodes, the command And re-execute the global model parameters issued by the federal server. Generate local model ; The process continues until the preset conditions are met, resulting in a pre-trained model.

Images