Intelligent consultation response system and method based on deep learning

Abstract

The invention relates to the technical field of deep learning, and discloses an intelligent consultation response system and method based on deep learning, and the system comprises a data acquisition module, a feature extraction module, a feature fusion module, a relevance calculation module, an infringement probability generation module, a clause screening module, and a reply generation module. The intellectual property consulting module is used for obtaining intellectual property consulting questions and legal texts of a user, generating a key clause set of the legal texts according to a multi-scale text convolution kernel and a multi-head attention mechanism, calculating semantic relevance between the intellectual property consulting questions and key clauses, constructing a knowledge graph of the key clauses, and displaying the knowledge graph. Performing multi-hop reasoning on the knowledge graph based on a graph neural network to obtain an infringement probability score, performing clause screening on the key clause set based on the infringement probability score and semantic relevance to obtain screened clauses, and generating a structured standard reply of the question according to the legal potency level of the screened clauses, and the response efficiency of the intellectual property consultation problem is improved.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology, and in particular to an intelligent consultation response system and method based on deep learning. Background Art

[0002] In the field of intellectual property, traditional consultation response systems usually rely on manual retrieval of legal provisions and precedents, which have problems such as low efficiency, high cost and limited coverage.

[0003] Traditional responses to IP consultations are mostly based on keyword matching or simple rule engines, which are difficult to capture the deep semantic associations of legal texts and can easily lead to omissions or mismatches of clauses. For example, complex legal texts have cross-chapter references, dynamic updates of judicial interpretations, and correlations with historical precedents. Traditional methods cannot effectively integrate multi-source heterogeneous data, resulting in delayed responses and insufficient accuracy.

[0004] In addition, the existing response methods lack the ability to conduct multi-dimensional quantitative analysis of infringement risks, and it is difficult to generate recommendations based on semantic similarity and legal effect levels to meet users' needs for efficient responses to intellectual property consultations. Summary of the invention

[0005] The present invention provides an intelligent consultation response system and method based on deep learning, the main purpose of which is to solve the problem of low response efficiency to intellectual property consultation issues.

[0006] To achieve the above-mentioned purpose, the present invention provides an intelligent consultation response system based on deep learning, characterized in that the system includes a data acquisition module, a feature extraction module, a feature fusion module, a correlation calculation module, an infringement probability generation module, a clause screening module and a reply generation module, wherein:

[0007] The data acquisition module is used to acquire the user's intellectual property consulting questions and associate them with the legal texts in the legal text database;

[0008] The feature extraction module is used to extract context features of the legal text based on the multi-scale text convolution kernel;

[0009] The feature fusion module is used to focus and fuse context features through a multi-head attention mechanism to obtain a set of key clauses of the legal text;

[0010] The relevance calculation module is used to calculate the semantic relevance between the intellectual property consulting questions and the key terms based on the contrastive learning algorithm;

[0011] The infringement probability generation module is used to construct a knowledge graph of key terms and perform multi-hop reasoning on the knowledge graph based on a graph neural network to obtain an infringement probability score for the key terms;

[0012] The clause screening module is used to mark high-risk clauses in the key clause set based on the infringement probability score, and to screen the key clause set in combination with semantic relevance to obtain screened clauses;

[0013] The response generation module is used to generate structured standard responses to intellectual property consulting questions based on the legal effect level of the screening clauses.

[0014] Optionally, when the data acquisition module acquires the user's intellectual property consulting questions and associates the legal texts in the legal text database, it includes:

[0015] Extracting legal entities in IP consulting issues;

[0016] Match legal entities with legal databases to obtain legal texts corresponding to intellectual property consulting issues.

[0017] Optionally, the method for constructing the multi-scale text convolution kernel in the feature extraction module includes:

[0018] Three kernel sizes are set to capture the semantics of short clauses, paragraphs, and cross-sections of legal texts;

[0019] A gating mechanism is used to suppress noise characteristics, where the gating function of the gating mechanism is:

[0020]

[0021] In the formula, is the context feature tensor, is the word vector of the legal text, is the bias term of multi-scale text convolution, is the convolution operation, is a weight matrix that adjusts the importance of different semantic features, yes function.

[0022] Optionally, when the feature fusion module performs focused fusion of context features through a multi-head attention mechanism to obtain a set of key clauses of the legal text, the feature fusion module includes:

[0023] Generate query matrix, key matrix and value matrix by linear transformation of context features;

[0024] Based on the query matrix, key matrix and value matrix, calculate the attention weights of the attention heads one by one;

[0025] The output features of the attention heads are concatenated based on the attention weights to generate a fused feature matrix;

[0026] The fusion feature matrix is ​​reduced in dimension through the fully connected layer to output the key terms feature vector;

[0027] Based on the L2 norm sorting of the key clause feature vectors, the first N vectors are retained as the key clause set of the legal text.

[0028] Optionally, the contrastive learning algorithm in the relevance calculation module is:

[0029]

[0030] In the formula, is the semantic relevance loss, is the embedding vector of the IP consulting question, is a negative sample, is the boundary threshold, is a positive sample, is the cosine similarity.

[0031] Optionally, when executing the construction of the knowledge graph of key terms, the infringement probability generation module includes:

[0032] Extract infringement elements of historical cases from the China Judgments Online;

[0033] Determine the nodes of the knowledge graph based on infringement elements, historical precedents and key clauses;

[0034] Mark the reference relationship of legal clauses and identify the correlation strength of infringement elements;

[0035] The edges of the knowledge graph are determined according to the citation relationship and the association strength, thereby generating a knowledge graph of key terms.

[0036] Optionally, the calculation formula of the infringement probability score in the infringement probability generation module is as follows:

[0037]

[0038] In the formula, is the infringement probability score, yes function, is to adjust the weight of the graph neural network output, is to aggregate the features of the current node and its neighboring nodes. is the current node feature, is the neighbor node feature, is the bias term in the graph neural network, is the identifier of the current node, Is the current node The set of neighbor nodes.

[0039] Optionally, the aggregation function used by the graph neural network in the infringement probability generation module is:

[0040]

[0041] in, It is Layer Current Node Features, is the rectified linear unit activation function, is the weight matrix that integrates the features of the current node and neighboring nodes, is a vector concatenation operation, It is Layer Current Node Features, is a mean operation, It is Layer Neighbor Node Features, is the identifier of the neighbor node, Is the current node The set of neighbor nodes of is the current number of layers of the graph neural network, is the identifier of the current node.

[0042] Optionally, the conflict resolution rules of the legal effect level in the reply generation module are:

[0043] The screening clauses are prioritized according to the legal effect level. The high-level clauses automatically override the conflicting content of the low-level clauses. The legal effect level is: laws prevail over regulations, and regulations prevail over precedents.

[0044] When a judicial interpretation conflicts with precedent, the most recently revised version of the judicial interpretation shall prevail.

[0045] In order to solve the above problems, the present invention also provides an intelligent consultation response method based on deep learning, the method comprising:

[0046] Obtain the user's intellectual property consulting questions and associate them with the legal texts in the legal text database;

[0047] Extracting contextual features of legal texts based on multi-scale text convolution kernels;

[0048] The context features are focused and integrated through a multi-head attention mechanism to obtain a set of key clauses in the legal text;

[0049] Calculate the semantic relevance between IP consulting questions and key clauses based on contrastive learning algorithm;

[0050] Construct a knowledge graph of key terms and perform multi-hop reasoning on the knowledge graph based on a graph neural network to obtain infringement probability scores for key terms;

[0051] High-risk clauses are marked in the key clause set based on the infringement probability score, and the key clause set is screened based on semantic relevance to obtain screened clauses;

[0052] Generate structured standard responses to IP consulting questions based on the legal effect level of the screening clauses.

[0053] Compared with the prior art, the present invention has the following beneficial effects:

[0054] 1. Use multi-scale text convolution kernels combined with a gating mechanism to accurately capture the local semantics and cross-chapter associations of legal texts and suppress noise interference;

[0055] 2. Dynamically focus on key terms through the multi-head attention mechanism, reduce redundant calculations, and improve processing speed;

[0056] 3. Calculate the semantic similarity between consulting questions and legal clauses based on contrastive learning algorithms, combine knowledge graphs with multi-hop reasoning of graph neural networks (GNNs), and achieve quantitative scoring of infringement probability to ensure the accuracy and explainability of clause recommendations;

[0057] 4. Automatically screen clauses based on the legal effectiveness level and the latest judicial interpretations, generate structured standard responses, avoid manual intervention, and reduce response time. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 A system architecture diagram of an intelligent consultation response system based on deep learning provided in one embodiment of the present invention;

[0059] Figure 2 A flowchart of a deep learning-based intelligent consultation response method provided in accordance with an embodiment of the present invention.

[0060] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0061] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments belong to a part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0062] The terms used in the embodiments of the present invention are only for the purpose of describing specific embodiments, and are not intended to limit the present invention. The singular forms of "said" and "the" used in the embodiments of the present invention and the appended claims are also intended to include plural forms, unless the context clearly indicates other meanings, and "multiple" generally includes at least two.

[0063] As used herein, the words “if” or “if” may be interpreted as “at the time of” or “when” or “in response to determining” or “in response to detecting”, depending on the context. Similarly, the phrases “if it is determined” or “if (stated condition or event) is detected” may be interpreted as “when it is determined” or “in response to determining” or “when detecting (stated condition or event)” or “in response to detecting (stated condition or event)”, depending on the context.

[0064] In addition, the step sequence in the following method embodiments is only an example and not a strict limitation.

[0065] In fact, the server-side device deployed by the intelligent consulting response system based on deep learning may be composed of one or more devices. The above-mentioned intelligent consulting response system based on deep learning can be implemented as: business instance, virtual machine, hardware device. For example, the intelligent consulting response system based on deep learning can be implemented as a business instance deployed on one or more devices in the cloud node. In simple terms, the intelligent consulting response system based on deep learning can be understood as a software deployed on the cloud node, which is used to provide each user end with an intelligent consulting response system based on deep learning. Alternatively, the intelligent consulting response system based on deep learning can also be implemented as a virtual machine deployed on one or more devices in the cloud node. The virtual machine is installed with application software for managing each user end. Alternatively, the intelligent consulting response system based on deep learning can also be implemented as a server consisting of many hardware devices of the same or different types, and one or more hardware devices are set to provide each user end with an intelligent consulting response system based on deep learning.

[0066] In terms of implementation, the intelligent consultation response system based on deep learning and the user end are adapted to each other. That is, the intelligent consultation response system based on deep learning is an application installed on the cloud service platform, and the user end is a client that establishes a communication connection with the application; or the intelligent consultation response system based on deep learning is implemented as a website, and the user end is implemented as a web page; or the intelligent consultation response system based on deep learning is implemented as a cloud service platform, and the user end is implemented as a small program in the instant messaging application.

[0067] like Figure 1 , is a system architecture diagram of a deep learning-based intelligent consultation response system provided in one embodiment of the present invention.

[0068] The intelligent consultation response system 100 based on deep learning of the present invention can be set in a cloud server. In terms of implementation, it can be used as one or more service devices, or as an application installed on the cloud (such as a server or server cluster of a mobile service operator), or it can be developed as a website. According to the functions implemented, the intelligent consultation response system 100 based on deep learning can include a data acquisition module 101, a feature extraction module 102, a feature fusion module 103, a correlation calculation module 104, an infringement probability generation module 105, a clause screening module 106 and a reply generation module 107. The module of the present invention can also be called a unit, which refers to a series of computer program segments that can be executed by a processor of an electronic device and can complete fixed functions, which are stored in the memory of the electronic device.

[0069] In the embodiment of the present invention, in the intelligent consulting response system based on deep learning, each of the above modules can be implemented independently and called with other modules. The call here can be understood as that a certain module can connect to multiple modules of another type and provide corresponding services to the multiple modules connected to it. In the intelligent consulting response system based on deep learning provided by the embodiment of the present invention, the scope of application of the intelligent consulting response system architecture based on deep learning can be adjusted by adding modules and directly calling them without modifying the program code, so as to achieve cluster-based horizontal expansion, so as to achieve the purpose of quickly and flexibly expanding the intelligent consulting response system based on deep learning. In practical applications, the above modules can be set in the same device or different devices, or they can be set in virtual devices, such as service instances in cloud servers.

[0070] In combination with specific embodiments, the following describes the various components and specific workflows of the intelligent consultation response system based on deep learning:

[0071] The data acquisition module is used to acquire the user's intellectual property consulting questions and associate them with the legal texts in the legal text database.

[0072] In the embodiment of the present invention, when the data acquisition module acquires the user's intellectual property consulting questions and associates the legal texts in the legal text database, it includes:

[0073] Extracting legal entities in IP consulting issues;

[0074] Match legal entities with legal databases to obtain legal texts corresponding to intellectual property consulting issues.

[0075] In detail, the data acquisition module receives the intellectual property consulting questions input by the user and associates the relevant clauses in the legal text database.

[0076] In detail, consulting questions are obtained through natural language interfaces (such as text input or speech-to-text), and entities are extracted: key entities in the questions (such as "patent number 2023XXXX") are identified through NLP technology, the search scope is narrowed, and the entities are mapped to legal terms (such as Article 26 of the Patent Law of the People's Republic of China) to ensure the accuracy of the associated text, reduce interference from irrelevant terms, and improve the processing efficiency of subsequent modules.

[0077] In detail, an embedding model (such as BERT) is used to vectorize the question, retrieve the most relevant clauses in the legal database (such as Article 26 of the Patent Law of the People’s Republic of China), and generate a preliminary set of related clauses.

[0078] In detail, the user inputs a consulting question (such as "Does unauthorized use of a trademark constitute infringement?"), and the "Trademark Law of the People's Republic of China", "Anti-Unfair Competition Law of the People's Republic of China" and related judicial interpretations are retrieved from the legal database as related texts.

[0079] The feature extraction module is used to extract context features of the legal text based on the multi-scale text convolution kernel.

[0080] In an embodiment of the present invention, the method for constructing a multi-scale text convolution kernel in the feature extraction module includes:

[0081] Three kernel sizes are set to capture the semantics of short clauses, paragraphs, and cross-sections of legal texts;

[0082] A gating mechanism is used to suppress noise characteristics, where the gating function of the gating mechanism is:

[0083]

[0084] In the formula, is the context feature tensor, is the word vector of the legal text, is the bias term of multi-scale text convolution, is the convolution operation, is a weight matrix that adjusts the importance of different semantic features, yes function.

[0085] In detail, the gating mechanism refers to the dynamic weighting through the Sigmoid function (such as retaining 80% of key features), suppressing redundant modifiers, and strengthening core terms (such as "technical features") to generate a gating value of 0-1, thereby improving the comprehensiveness and noise resistance of legal text feature extraction.

[0086] In detail, three types of convolution kernels (width 3 / 5 / 7) are used to extract local features of legal texts, capturing short clauses, paragraphs, and cross-chapter semantics respectively. For example, convolution kernel width 3 captures phrase-level features (such as "unauthorized use"), width 5 captures long-distance dependencies (such as "serious circumstances require criminal liability"). A 7×7 kernel is used across chapters to identify the cross-chapter association between "patent infringement" and "compensation calculation".

[0087] In detail, the context feature tensor represents the features of the legal text after being filtered by the gating mechanism, which is used for subsequent semantic analysis; the word vector of the legal text can be a 300-dimensional Word2Vec or BERT embedding; the convolution operation refers to the extraction of local semantic features through multi-scale convolution kernels (such as 2×300, 3×300, 5×300); the bias term is used to enhance the nonlinear expression ability of the model; the Sigmoid function ensures that the output value is between 0-1, which is used to control the retention ratio of the convolution result (such as 0.8 means retaining 80% of the features).

[0088] In detail, the gating mechanism suppresses noise in legal texts (such as redundant modifiers) through dynamic weight allocation, while strengthening the characteristic expressions of key terms (such as "similar trademarks" and "technical features") to improve the accuracy of subsequent clause matching.

[0089] The feature fusion module is used to focus and fuse context features through a multi-head attention mechanism to obtain a set of key clauses of the legal text.

[0090] In the embodiment of the present invention, when the feature fusion module performs focused fusion of context features through a multi-head attention mechanism to obtain a set of key clauses of a legal text, the feature fusion module includes:

[0091] Generate query matrix, key matrix and value matrix by linear transformation of context features;

[0092] Based on the query matrix, key matrix and value matrix, calculate the attention weights of the attention heads one by one;

[0093] The output features of the attention heads are concatenated based on the attention weights to generate a fused feature matrix;

[0094] The fusion feature matrix is ​​reduced in dimension through the fully connected layer to output the key terms feature vector;

[0095] Based on the L2 norm sorting of the key clause feature vectors, the first N vectors are retained as the key clause set of the legal text.

[0096] In detail, the multi-head attention mechanism uses 8 parallel attention heads to capture the semantic relations of different subspaces (such as the effectiveness of clauses and infringement scenarios), and the attention weights are normalized by softmax and then weighted summed with multi-scale features to focus on the core content of the legal text and avoid information overload.

[0097] Specifically, the top N high-weight terms (e.g., Top 20%) are retained to ensure the saliency and interpretability of the output set.

[0098] In detail, a multi-head attention mechanism is used to dynamically focus on the local and global semantics of the legal text to solve the problem that traditional methods cannot capture the correlation between complex clauses. L2 norm sorting is combined to screen key clauses to ensure the significance and interpretability of the output set.

[0099] In general, the gating mechanism filters out noise and retains the core features of the legal text (such as "similar trademarks" and "technical features"); the multi-scale convolution kernel captures the details of the clauses (short sentences), paragraph logic (judicial interpretation) and cross-text associations (legal and case citations).

[0100] That is, through the gating mechanism and multi-scale convolution, key clauses can be located from massive legal texts.

[0101] The relevance calculation module is used to calculate the semantic relevance between intellectual property consulting questions and key terms based on a contrastive learning algorithm.

[0102] In the embodiment of the present invention, the contrastive learning algorithm in the association calculation module is:

[0103]

[0104] In the formula, is the semantic relevance loss, is the embedding vector of the IP consulting question, is a negative sample, is the boundary threshold, is a positive sample, is the cosine similarity.

[0105] In detail, positive samples are relevant clauses, such as Article 10 of the Copyright Law of the People's Republic of China; negative samples are randomly sampled irrelevant clauses; the boundary threshold is set to 0.2 to ensure that the similarity of positive samples is at least 0.2 higher than that of negative samples; the relevance score matrix is ​​composed of question-clause similarities.

[0106] In detail, the contrastive learning algorithm completes the semantic relevance calculation by maximizing the similarity between the consulting question and the positive sample legal clauses and minimizing its similarity with the negative sample clauses.

[0107] That is, construct positive samples (real relevant terms corresponding to consulting questions) and negative samples (randomly sampled irrelevant terms), encode the questions and terms texts through the BERT model, and calculate the cosine similarity as the relevance score.

[0108] In detail, the semantic relevance loss measures the matching difference between the consulting question and the positive and negative sample clauses, and is used to optimize the model; the positive samples are legal clauses that are semantically related to the consulting question (such as "trademark infringement" corresponds to Article 57 of the Trademark Law of the People's Republic of China); the negative samples are randomly sampled irrelevant clauses (such as Article 10 of the Copyright Law of the People's Republic of China); the boundary threshold (usually set to 0.2) forces the positive sample similarity to be at least 0.2 higher than the negative sample to avoid the model confusing easily confused categories; the cosine similarity is used to quantify the similarity of two semantic vectors.

[0109] In detail, through comparative learning, the model can better distinguish between easily confused legal issues such as "patent infringement" and "trademark infringement", ensuring that the recommended terms are highly relevant to user needs.

[0110] In detail, contrastive learning forces the distinction between positive and negative samples to ensure that the recommended terms are highly relevant to user questions. For example, the similarity of positive samples (relevant terms) is forced to be at least 0.2 (such as 0.8 vs 0.5) higher than that of negative samples (irrelevant terms) to avoid confusion; cosine similarity quantifies semantic relevance (such as "trademark similarity" and Article 57 of the Trademark Law of the People's Republic of China), replacing traditional keyword matching, using contrastive learning and cosine similarity to ensure that the recommended content is highly consistent with user needs.

[0111] The infringement probability generation module is used to construct a knowledge graph of key terms and perform multi-hop reasoning on the knowledge graph based on a graph neural network to obtain an infringement probability score for the key terms.

[0112] In the embodiment of the present invention, when executing the construction of the knowledge graph of key terms, the infringement probability generation module includes:

[0113] Extract infringement elements of historical cases from the China Judgments Online;

[0114] Determine the nodes of the knowledge graph based on infringement elements, historical precedents and key clauses;

[0115] Mark the reference relationship of legal clauses and identify the correlation strength of infringement elements;

[0116] The edges of the knowledge graph are determined according to the citation relationship and the association strength, thereby generating a knowledge graph of key terms.

[0117] In detail, Stanford CoreNLP is used to automatically annotate legal clause citation relationships.

[0118] In detail, the nodes of the knowledge graph represent precedents, legal clauses, and infringement elements (technical feature overlap, trademark similarity); the edges of the knowledge graph represent the precedent-clause citation relationship and the element association strength.

[0119] Furthermore, the knowledge graph integrates legal provisions (such as Article 11 of the Patent Law of the People's Republic of China), historical precedents (from the China Judgments Online website), and infringement factors (such as "overlap of technical features"); annotates citation relationships (such as case A citing clause B) and correlation strength (co-occurrence frequency > 5 times is a strong correlation), providing structured data support for graph neural networks and enhancing the credibility of infringement probability reasoning.

[0120] In detail, the multi-hop reasoning aggregates the semantic information of nodes within a 3-hop range through the neighborhood message passing mechanism of the graph neural network, and calculates the infringement probability score of the clause node.

[0121] In detail, the legal clause nodes are embedded with the BERT vector of the clause text, and the historical case nodes are embedded with the case description; the edges include "clause A is cited by case B" and "case C and case D involve the same type of infringement."

[0122] In the embodiment of the present invention, the calculation formula of the infringement probability score in the infringement probability generation module is as follows:

[0123]

[0124] In the formula, is the infringement probability score, yes function, is to adjust the weight of the graph neural network output, is to aggregate the features of the current node and its neighboring nodes. is the current node feature, is the neighbor node feature, is the bias term in the graph neural network, is the identifier of the current node, Is the current node The set of neighbor nodes.

[0125] In detail, the infringement probability score indicates the possibility that a certain clause will be cited in infringement precedents, and the higher the value, the greater the risk; the Sigmoid function maps the linear output to a probability value; the current node feature represents the semantic vector (e.g., 256 dimensions) of the legal clause or precedent; the neighbor node feature set includes directly related precedents, infringement elements, etc.; the bias term is used to enhance the flexibility of the model.

[0126] In detail, through multi-hop reasoning of graph neural networks (such as analyzing the relationship between "overlapping technical features" and multiple precedents), the infringement risk of a certain clause can be predicted, providing a quantitative basis for legal advice.

[0127] In the embodiment of the present invention, the aggregation function used by the graph neural network in the infringement probability generation module is:

[0128]

[0129] in, It is Layer Current Node Features, is the rectified linear unit activation function, is the weight matrix that integrates the features of the current node and neighboring nodes, is a vector concatenation operation, It is Layer Current Node Features, is a mean operation, It is Layer Neighbor Node Features, is the identifier of the neighbor node, Is the current node The set of neighbor nodes of is the current number of layers of the graph neural network, is the identifier of the current node.

[0130] In detail, the information of neighboring nodes is aggregated through the graph neural network (GNN) to calculate the infringement probability; the Sigmoid function maps the GNN output to a probability value (such as 0.75 indicates high risk); the mean operation fuses the characteristics of neighboring nodes (such as the losing rate of related cases), and the nonlinear expression ability is enhanced through ReLU activation to quantify the infringement risk of the clauses and support multi-dimensional legal advice (such as "high-risk clauses need to be avoided").

[0131] In detail, is the output of the previous layer; the mean operation refers to averaging the features of neighboring nodes to reduce the impact of noise; Usually set to 2 layers to balance model depth and overfitting risk; node The set of neighbors can be all cases that cite the same legal provision.

[0132] In detail, by aggregating the features of neighbor nodes (such as the loss rate in case law and the comparison results of technical features), the contextual relevance of legal provisions can be captured, and the accuracy of infringement probability prediction can be improved.

[0133] In detail, the knowledge graph integrates legal clauses, precedents and infringement elements to construct a multi-dimensional association network; GNN aggregates neighbor node information (such as loss rate, technology overlap) and outputs an infringement probability score.

[0134] The clause screening module is used to mark high-risk clauses in the key clause set based on the infringement probability score, and to screen the key clause set in combination with semantic relevance to obtain screened clauses.

[0135] Specifically, if a clause is cited in historical cases and has a losing rate greater than 70%, it will be marked as “high risk”.

[0136] In an embodiment of the present invention, when the clause screening module performs clause screening on the key clause set in combination with semantic relevance to obtain the screening clauses, it includes: screening the final clauses in combination with the infringement probability and semantic relevance, wherein the clauses with a semantic similarity greater than 0.65 are retained. For example: when the infringement probability of clause A is 75% and the similarity is 0.8, clause A is retained; when the infringement probability of clause B is 65% and the similarity is 0.7, clause A is eliminated.

[0137] The response generation module is used to generate structured standard responses to intellectual property consulting questions based on the legal effect level of the screening clauses.

[0138] In the embodiment of the present invention, the conflict resolution rules of the legal effect level in the reply generation module are:

[0139] The screening clauses are prioritized according to the legal effect level. The high-level clauses automatically override the conflicting content of the low-level clauses. The legal effect level is: laws prevail over regulations, and regulations prevail over precedents.

[0140] When a judicial interpretation conflicts with precedent, the most recently revised version of the judicial interpretation shall prevail.

[0141] In detail, the recommendations are sorted by the level of legal effect (laws>regulations>precedents) to ensure the authority of the recommendations, with higher levels automatically covering lower-level conflicting content (such as the Trademark Law of the People's Republic of China takes precedence over local regulations). When judicial interpretations conflict with precedents, the latest revised version of the judicial interpretation is given priority (such as the 2023 revision takes precedence over the 2020 revision), so as to resolve conflicts in terms, avoid contradictory legal advice, and enhance user trust.

[0142] Reference Figure 2 FIG. 1 is a flow chart of an intelligent consultation response method based on deep learning provided by an embodiment of the present invention. In this embodiment, the intelligent consultation response method based on deep learning includes:

[0143] S1. Obtain the user's intellectual property consulting questions and associate them with the legal texts in the legal text database;

[0144] S2, extracting contextual features of legal texts based on multi-scale text convolution kernels;

[0145] S3, focusing on and fusing context features through a multi-head attention mechanism to obtain a set of key clauses in the legal text;

[0146] S4. Calculate the semantic relevance between IP consulting questions and key clauses based on contrastive learning algorithm;

[0147] S5. Build a knowledge graph of key terms and perform multi-hop reasoning on the knowledge graph based on a graph neural network to obtain infringement probability scores for key terms;

[0148] S6. High-risk clauses are marked in the key clause set based on the infringement probability score, and the key clause set is screened based on semantic relevance to obtain screened clauses;

[0149] S7. Generate structured standard responses to intellectual property consulting questions based on the legal effect level of the screening clauses.

[0150] It is obvious to those skilled in the art that the present invention is not limited to the details of the above exemplary embodiments, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0151] The embodiments of the present application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence is the theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results.

[0152] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the spirit and scope of the technical solution of the present invention.

Claims

1. An intelligent consultation response system based on deep learning, characterized in that: The system includes a data acquisition module, a feature extraction module, a feature fusion module, a correlation calculation module, an infringement probability generation module, a clause screening module and a response generation module, wherein: The data acquisition module is used to acquire the user's intellectual property consulting questions and associate them with the legal texts in the legal text database; The feature extraction module is used to extract context features of the legal text based on the multi-scale text convolution kernel; The feature fusion module is used to focus and fuse context features through a multi-head attention mechanism to obtain a set of key clauses of the legal text; The relevance calculation module is used to calculate the semantic relevance between the intellectual property consulting questions and the key terms based on the contrastive learning algorithm; The infringement probability generation module is used to construct a knowledge graph of key terms and perform multi-hop reasoning on the knowledge graph based on a graph neural network to obtain an infringement probability score for the key terms; The clause screening module is used to mark high-risk clauses in the key clause set based on the infringement probability score, and screen the key clause set in combination with semantic relevance to obtain screened clauses; The response generation module is used to generate structured standard responses to intellectual property consulting questions based on the legal effect level of the screening clauses.

2. The deep learning-based intelligent consultation response system according to claim 1, characterized in that: When the data acquisition module acquires the user's intellectual property consulting questions and associates the legal texts in the legal text database, it includes: Extracting legal entities in IP consulting issues; Match legal entities with legal databases to obtain legal texts corresponding to intellectual property consulting issues.

3. The deep learning-based intelligent consultation response system according to claim 1, characterized in that: The method for constructing the multi-scale text convolution kernel in the feature extraction module includes: Three kernel sizes are set to capture the semantics of short clauses, paragraphs, and cross-sections of legal texts; A gating mechanism is used to suppress noise characteristics, where the gating function of the gating mechanism is: In the formula, is the context feature tensor, is the word vector of the legal text, is the bias term of multi-scale text convolution, is the convolution operation, is a weight matrix that adjusts the importance of different semantic features, yes function.

4. The deep learning-based intelligent consultation response system according to claim 1, characterized in that: The feature fusion module performs focused fusion of context features through a multi-head attention mechanism to obtain a set of key clauses of the legal text, including: Generate query matrix, key matrix and value matrix by linear transformation of context features; Based on the query matrix, key matrix and value matrix, calculate the attention weights of the attention heads one by one; The output features of the attention heads are concatenated based on the attention weights to generate a fused feature matrix; The fusion feature matrix is ​​reduced in dimension through the fully connected layer to output the key terms feature vector; Based on the L2 norm sorting of the key clause feature vectors, the first N vectors are retained as the key clause set of the legal text.

5. The deep learning-based intelligent consultation response system according to claim 1, characterized in that: The contrastive learning algorithm in the association calculation module is: In the formula, is the semantic relevance loss, is the embedding vector of the IP consulting question, is a negative sample, is the boundary threshold, is a positive sample, is the cosine similarity.

6. The deep learning-based intelligent consultation response system according to claim 1, characterized in that: When executing the construction of the knowledge graph of key terms, the infringement probability generation module includes: Extract infringement elements of historical cases from the China Judgments Online; Determine the nodes of the knowledge graph based on infringement elements, historical precedents and key clauses; Mark the reference relationship of legal clauses and identify the correlation strength of infringement elements; The edges of the knowledge graph are determined according to the citation relationship and the association strength, thereby generating a knowledge graph of key terms.

7. The deep learning-based intelligent consultation response system according to claim 1, characterized in that: The calculation formula of the infringement probability score in the infringement probability generation module is as follows: In the formula, is the infringement probability score, yes function, is to adjust the weight of the graph neural network output, is to aggregate the features of the current node and its neighboring nodes. is the current node feature, is the neighbor node feature, is the bias term in the graph neural network, is the identifier of the current node, Is the current node The set of neighbor nodes.

8. The deep learning-based intelligent consultation response system according to claim 1, characterized in that: The aggregation function used by the graph neural network in the infringement probability generation module is: in, It is Layer Current Node Features, is the rectified linear unit activation function, is the weight matrix that integrates the features of the current node and neighboring nodes, is a vector concatenation operation, It is Layer Current Node Features, is a mean operation, It is Layer Neighbor Node Features, is the identifier of the neighbor node, Is the current node The set of neighbor nodes of is the current number of layers of the graph neural network, is the identifier of the current node.

9. The deep learning-based intelligent consultation response system according to any one of claims 1 to 8, characterized in that: The conflict resolution rules at the legal effect level in the reply generation module are: The screening clauses are prioritized according to the legal effect level. The high-level clauses automatically override the conflicting content of the low-level clauses. The legal effect level is: laws prevail over regulations, and regulations prevail over precedents. When a judicial interpretation conflicts with precedent, the most recently revised version of the judicial interpretation shall prevail.

10. An intelligent consultation response method based on deep learning, characterized in that: The method comprises: S1. Obtain the user's intellectual property consulting questions and associate them with the legal texts in the legal text database; S2, extracting contextual features of legal texts based on multi-scale text convolution kernels; S3, focusing on and fusing context features through a multi-head attention mechanism to obtain a set of key clauses in the legal text; S4. Calculate the semantic relevance between IP consulting questions and key clauses based on contrastive learning algorithm; S5. Build a knowledge graph of key terms and perform multi-hop reasoning on the knowledge graph based on a graph neural network to obtain infringement probability scores for key terms; S6. High-risk clauses are marked in the key clause set based on the infringement probability score, and the key clause set is screened based on semantic relevance to obtain screened clauses; S7. Generate structured standard responses to intellectual property consulting questions based on the legal effect level of the screening clauses.

Images