An intelligent consultation response system and method based on deep learning

Abstract

The present invention relates to the technical field of deep learning, and discloses an intelligent consultation response system and method based on deep learning. The system includes 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 response generation module, which are used to obtain the user's intellectual property consultation questions and legal texts, generate a set of key clauses of the legal text according to multi-scale text convolutional kernels and multi-head attention mechanisms, calculate the semantic relevance between the intellectual property consultation questions and the key clauses, construct a knowledge graph of the key clauses, perform multi-hop reasoning on the knowledge graph based on a graph neural network to obtain an infringement probability score, screen the set of key clauses based on the infringement probability score and semantic relevance to obtain screened clauses, and generate a structured standard response to the question according to the legal effect level of the screened clauses, thereby improving the response efficiency of intellectual property consultation questions.

Description

Technical Field

[0001] The present invention relates to the technical field of deep learning, 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 case precedents, which have problems such as low efficiency, high cost, and limited coverage.

[0003] Traditional intellectual property consultation response methods are mostly based on keyword matching or simple rule engines, which are difficult to capture the deep semantic associations of legal texts and are prone to clause omissions or mis-matches. For example, there are cross-chapter references, dynamic updates of judicial interpretations, and correlations of historical case precedents among complex legal provisions. Traditional methods cannot effectively integrate multi-source heterogeneous data, resulting in response delays and insufficient accuracy.

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

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

[0006] To achieve the above object, an intelligent consultation response system based on deep learning provided by the present invention is characterized in that the system includes 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, wherein:

[0007] The data acquisition module is used to acquire the intellectual property consultation problem of the user and associate the legal texts in the legal text database;

[0008] The feature extraction module is used to extract the context features of legal texts according to multi-scale text convolutional kernels;

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

[0010] The relevance calculation module is used to calculate the semantic relevance between the intellectual property consultation problem and the key clauses based on a contrast learning algorithm;

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

[0012] 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;

[0013] The reply generation module is used to generate a structured standard reply to the intellectual property consultation question according to the legal effect level of the screened clauses.

[0014] Optionally, when the data acquisition module executes to obtain the intellectual property consultation question of the user and associate the legal text in the legal text database, it includes:

[0015] Extract the legal entities in the intellectual property consultation question;

[0016] Match the legal entities with the legal database to obtain the legal text corresponding to the intellectual property consultation question.

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

[0018] Set 3 kernel sizes, which are respectively used to capture short clauses, paragraphs and cross-chapter semantics of legal texts;

[0019] Adopt a gating mechanism to suppress noise features, where the gating function of the gating mechanism is:

[0020] Output=σ(W g ·x + b g )·Conv(x)

[0021] In the formula, Output is the context feature tensor, x is the word vector of the legal text, b g is the bias term of the multi-scale text convolution, Conv(x) is the convolution operation, W g is the weight matrix that adjusts the importance of different semantic features, and σ(*) is the Sigmoid function.

[0022] Optionally, when the feature fusion module executes to focus and fuse the context features through the multi-head attention mechanism to obtain the key clause set of the legal text, it includes:

[0023] Generate a query matrix, a key matrix and a value matrix by linearly transforming the context features;

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

[0025] Based on the attention weights, splice the output features of the attention heads to generate a fusion feature matrix;

[0026] Reduce the dimension of the fused feature matrix through a fully connected layer and output the key clause feature vector;

[0027] Based on the L2 norm sorting of the key clause feature vectors, retain the top N vectors as the set of key clauses of the legal text.

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

[0029] L = max(0, sim(q, p - ) - sim(q, p + ) + γ)

[0030] where L is the semantic relevance loss, q is the embedding vector of the intellectual property consultation question, p - is the negative sample, γ is the margin threshold, p + is the positive sample, and sim(*) is the cosine similarity.

[0031] Optionally, when the infringement probability generation module executes the construction of the knowledge graph of key clauses, it includes:

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

[0033] Determine the nodes of the knowledge graph according to the infringement elements, historical cases and key clauses;

[0034] Mark the legal clause citation relationship and identify the association strength of the infringement elements;

[0035] Determine the edges of the knowledge graph according to the citation relationship and the association strength, thereby generating the knowledge graph of key clauses.

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

[0037]

[0038] where P 侵权 is the infringement probability score, σ(*) is the Sigmoid function, W a is the weight for adjusting the output of the graph neural network, is to aggregate the current node feature and its neighbor node features, h v is the current node feature, is the neighbor node feature, b is the bias term in the graph neural network, v is the identifier of the current node, is the set of neighbor nodes of the current node v.

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

[0040]

[0041] Among them, is the feature of the current node v at the k-th layer, ReLU(*) is the rectified linear unit activation function, and W b is the weight matrix for fusing the features of the current node and its neighbor nodes, and CONCAT(*) is the vector concatenation operation. is the feature of the current node v at the (k - 1)-th layer, and MEAN(*) is the mean operation. is the feature of the neighbor node u at the (k - 1)-th layer, where u is the identifier of the neighbor node. is the set of neighbor nodes of the current node v, k is the current layer number of the graph neural network, and v is the identifier of the current node.

[0042] Optionally, the conflict resolution rule for the legal effect levels in the reply generation module is as follows:

[0043] Sort the screening clauses according to the legal effect levels, and the clauses at the higher level automatically override the conflicting content of the clauses at the lower level. Among them, the legal effect levels are: laws take precedence over regulations, and regulations take precedence over judicial precedents.

[0044] When there is a conflict between judicial interpretations and judicial precedents, the latest revised version of the judicial interpretation is adopted.

[0045] To solve the above problems, the present invention also provides an intelligent consultation response method based on deep learning. The method includes:

[0046] Obtain the intellectual property consultation question of the user and associate the legal texts in the legal text database.

[0047] Extract the context features of the legal texts according to the multi-scale text convolutional kernels.

[0048] Focus and fuse the context features through the multi-head attention mechanism to obtain the set of key clauses of the legal texts.

[0049] Calculate the semantic relevance between the intellectual property consultation question and the key clauses based on the contrast learning algorithm.

[0050] Construct a knowledge graph of the key clauses and perform multi-hop reasoning on the knowledge graph based on the graph neural network to obtain the infringement probability score of the key clauses.

[0051] Mark the high-risk clauses in the set of key clauses based on the infringement probability score, and combine the semantic relevance to screen the clauses in the set of key clauses to obtain the screening clauses.

[0052] Generate a structured standard reply to the intellectual property consultation question according to the legal effect levels of the screening clauses. Compared with the prior art, the present invention has the following beneficial effects:

[0053] 1. Adopt a multi-scale text convolution kernel combined with a gating mechanism to accurately capture the local semantics and cross-chapter associations of legal texts and suppress noise interference;

[0054] 2. Dynamically focus on key clauses through the multi-head attention mechanism to reduce redundant calculations and improve processing speed;

[0055] 3. Calculate the semantic similarity between consulting questions and legal clauses based on the contrast learning algorithm, and combine the multi-hop reasoning of the knowledge graph and the graph neural network (GNN) to achieve a quantitative score of the infringement probability and ensure the accuracy and interpretability of clause recommendations;

[0056] 4. Automatically screen clauses according to the legal effect level and the latest judicial interpretations, generate a structured standard response, avoid manual intervention, and reduce the response time. Brief Description of the Drawings

[0057] Figure 1 It is a system architecture diagram of an intelligent consulting response system based on deep learning provided by an embodiment of the present invention;

[0058] Figure 2 It is a schematic flowchart of an intelligent consulting response method based on deep learning provided by an embodiment of the present invention.

[0059] The realization, functional features, and advantages of the purpose of the present invention will be further described in conjunction with the embodiments with reference to the drawings. Detailed Embodiments

[0060] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions 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 are part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0061] 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 "the" and "said" used in the embodiments of the present invention and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. "A plurality" generally includes at least two.

[0062] Depending on the context, as used herein, the words "if" or "when" may be interpreted as "when...", "while...", "in response to determining", or "in response to detecting". Similarly, depending on the context, the phrase "if determined" or "if detecting (stated condition or event)" may be interpreted as "when determined", "in response to determining", "when detecting (stated condition or event)", or "in response to detecting (stated condition or event)".

[0063] In addition, the step timings in the following method embodiments are only examples, not strictly limited.

[0064] In fact, the server devices deployed in the intelligent consultation response system based on deep learning may be composed of one or more devices. The above-mentioned intelligent consultation response system based on deep learning can be implemented as: a business instance, a virtual machine, or a hardware device. For example, the intelligent consultation response system based on deep learning can be implemented as a business instance deployed on one or more devices in a cloud node. Briefly, the intelligent consultation response system based on deep learning can be understood as a software deployed on a cloud node for providing the intelligent consultation response system based on deep learning to each client. Or, the intelligent consultation response system based on deep learning can also be implemented as a virtual machine deployed on one or more devices in a cloud node. An application software for managing each client is installed in the virtual machine. Or, the intelligent consultation response system based on deep learning can also be implemented as a server composed of many identical or different types of hardware devices, and one or more hardware devices are set to provide the intelligent consultation response system based on deep learning to each client.

[0065] In terms of implementation form, the intelligent consultation response system based on deep learning and the client adapt to each other. That is, if the intelligent consultation response system based on deep learning is an application installed on a cloud service platform, then the client is a client that establishes a communication connection with the application; or if the intelligent consultation response system based on deep learning is implemented as a website, then the client is implemented as a web page; or if the intelligent consultation response system based on deep learning is implemented as a cloud service platform, then the client is implemented as a small program in an instant messaging application.

[0066] As Figure 1 shown, it is the system architecture diagram of the intelligent consultation response system based on deep learning provided by an embodiment of the present invention.

[0067] The intelligent consultation response system 100 based on deep learning according to the present invention can be set in a cloud server. In terms of implementation form, it can be used as one or more service devices, or can be installed as an application on the cloud (such as the server of a mobile service operator, a server cluster, etc.), or can also be developed into a website. According to the functions to be realized, the intelligent consultation response system 100 based on deep learning may include a data acquisition module 101, a feature extraction module 102, a feature fusion module 103, a relevance calculation module 104, an infringement probability generation module 105, a clause screening module 106, and a response generation module 107. The modules in the present invention can also be referred to as units, which refer to a series of computer program segments that can be executed by a processor of an electronic device and can complete fixed functions, and are stored in the memory of the electronic device.

[0068] In an embodiment of the present invention, in the intelligent consultation response system based on deep learning, each of the above modules can be independently implemented and called with other modules. Here, the call can be understood as that a certain module can be connected to multiple modules of another type and provide corresponding services for the multiple modules it is connected to. In the intelligent consultation response system provided by the embodiment of the present invention, without modifying the program code, the applicable range of the architecture of the intelligent consultation response system based on deep learning can be adjusted by adding modules and directly calling them, so as to achieve cluster-level horizontal expansion, so as to achieve the purpose of quickly and flexibly expanding the intelligent consultation response system based on deep learning. In practical applications, the above modules can be set in the same device or different devices, or can also be set in virtual devices, such as service instances in a cloud server.

[0069] The following will respectively describe each component and the specific working process of the intelligent consultation response system based on deep learning in combination with specific embodiments:

[0070] The data acquisition module is used to obtain the intellectual property consultation questions of users and associate the legal texts in the legal text database.

[0071] In an embodiment of the present invention, when the data acquisition module executes to obtain the intellectual property consultation questions of users and associate the legal texts in the legal text database, it includes:

[0072] Extract the legal entities in the intellectual property consultation questions;

[0073] Match the legal entities with the legal database to obtain the legal texts corresponding to the intellectual property consultation questions.

[0074] Specifically, the data acquisition module receives the intellectual property consultation questions input by users and associates the relevant clauses in the legal text database.

[0075] Specifically, the consultation question is obtained through a natural language interface (such as text input or speech-to-text), and entities are extracted: the key entities in the question are identified through NLP technology to narrow the retrieval scope, map the entities to legal provisions, ensure the accuracy of the associated text, reduce interference from irrelevant provisions, and improve the processing efficiency of subsequent modules.

[0076] Specifically, an embedding model (such as BERT) is used to vectorize the question, and the most relevant provisions in the legal database are retrieved to generate a set of initially associated provisions.

[0077] Specifically, the user inputs a consultation question (such as "Does the unauthorized use of a trademark constitute infringement?"), and relevant text is retrieved from the legal database as the associated text.

[0078] The feature extraction module is used to extract the context features of legal texts according to multi-scale text convolution kernels.

[0079] In the embodiment of the present invention, the construction method of the multi-scale text convolution kernel in the feature extraction module includes:

[0080] Set 3 kernel sizes, which are respectively used to capture short clauses, paragraphs, and cross-chapter semantics of legal texts;

[0081] Adopt a gating mechanism to suppress noise features, where the gating function of the gating mechanism is:

[0082] Output = σ(W g ·x + b g )·Conv(x)

[0083] In the formula, Output is the context feature tensor, x is the word vector of the legal text, b g is the bias term of the multi-scale text convolution, Conv(x) is the convolution operation, W g is the weight matrix that adjusts the importance of different semantic features, and σ(*) is the Sigmoid function.

[0084] Specifically, the gating mechanism refers to dynamically weighting through the Sigmoid function (such as retaining 80% of the key features), suppressing redundant modifiers, strengthening core terms (such as "technical features"), and generating a gating value of 0-1 to improve the comprehensiveness and noise resistance of legal text feature extraction.

[0085] Specifically, 3 convolution kernels (width 3 / 5 / 7) are used to extract the local features of legal texts, capturing short clauses, paragraphs, and cross-chapter semantics respectively. For example, a convolution kernel width of 3 captures phrase-level features (such as "use without permission"), width 5 captures long-distance dependencies (such as "those with serious circumstances shall bear criminal liability"), and a 7×7 kernel is used for cross-chapter to identify the cross-chapter association between "patent infringement" and "compensation calculation".

[0086] Specifically, the context feature tensor represents the legal text features screened by the gating mechanism and is used for subsequent semantic analysis; the word vectors of the legal text can be 300-dimensional Word2Vec or BERT embeddings; the convolution operation refers to extracting local semantic features through multi-scale convolutional kernels (such as 2×300, 3×300, 5×300); the bias term is used to enhance the non-linear expression ability of the model; the Sigmoid function ensures that the output value is between 0 and 1 and is used to control the retention ratio of the convolution result (for example, 0.8 means retaining 80% of the features).

[0087] Specifically, the gating mechanism suppresses the noise in the legal text (such as redundant modifiers) through dynamic weight allocation, while strengthening the feature expression of key terms (such as "similar trademarks" and "technical features"), and improving the accuracy of subsequent clause matching.

[0088] The feature fusion module is used to focus and fuse the context features through the multi-head attention mechanism to obtain the key clause set of the legal text.

[0089] In the embodiment of the present invention, when the feature fusion module executes the focusing and fusion of the context features through the multi-head attention mechanism to obtain the key clause set of the legal text, it includes:

[0090] Generating a query matrix, a key matrix, and a value matrix by linearly transforming the context features;

[0091] Based on the query matrix, the key matrix, and the value matrix, calculating the attention weights of each attention head one by one;

[0092] Based on the attention weights, splicing the output features of the attention heads to generate a fusion feature matrix;

[0093] Reducing the dimension of the fusion feature matrix through a fully connected layer and outputting the key clause feature vector;

[0094] Based on the L2 norm sorting of the key clause feature vectors, retaining the top N vectors as the key clause set of the legal text.

[0095] Specifically, the multi-head attention mechanism uses 8 parallel attention heads to capture the semantic relationships in different subspaces (such as clause effectiveness and infringement scenarios), and the attention weights are weighted and summed with multi-scale features after being normalized by softmax to focus on the core content of the legal text and avoid information overload.

[0096] Specifically, retaining the top N high-weight clauses (such as Top 20%) to ensure the significance and interpretability of the output set.

[0097] Specifically, the multi-head attention mechanism is used to dynamically focus on the local and global semantics of legal texts, solving the problem that traditional methods are insufficient in capturing the relevance of complex clauses; the L2 norm sorting is combined to screen key clauses, ensuring the significance and interpretability of the output set.

[0098] Generally speaking, the gating mechanism filters out noise and retains the core features of legal texts (such as "similar trademarks", "technical features"); the multi-scale convolutional kernels respectively capture clause details (short sentences), paragraph logic (judicial interpretations), and cross-text associations (legal and case citations).

[0099] That is, through the gating mechanism and multi-scale convolution, key clauses are located from a large number of legal texts.

[0100] The relevance calculation module is used to calculate the semantic relevance between the intellectual property consultation question and the key clause based on the contrast learning algorithm.

[0101] In the embodiment of the present invention, the contrast learning algorithm in the relevance calculation module is:

[0102] L = max(0, sim(q, p - ) - sim(q, p + ) + γ)

[0103] In the formula, L is the semantic relevance loss, q is the embedding vector of the intellectual property consultation question, p - is the negative sample, γ is the margin threshold, p + is the positive sample, and sim(*) is the cosine similarity.

[0104] Specifically, the positive sample is the relevant clause; the negative sample is the randomly sampled irrelevant clause; the margin threshold is set to 0.2 to ensure that the similarity of the positive sample is at least 0.2 higher than that of the negative sample; the relevance score matrix is composed of the question-clause similarities.

[0105] Specifically, the contrast learning algorithm completes the semantic relevance calculation by maximizing the similarity between the consultation question and the positive sample legal clause and minimizing its similarity with the negative sample clause.

[0106] That is, positive samples (true relevant clauses corresponding to the consultation question) and negative samples (randomly sampled irrelevant clauses) are constructed, and the question and clause texts are encoded by the BERT model, and the cosine similarity is calculated as the relevance score.

[0107] Specifically, the semantic relevance loss measures the matching differences between the consulting question and the positive and negative sample clauses, which is used to optimize the model; the positive samples are legal clauses semantically related to the consulting question; the negative samples are randomly sampled irrelevant clauses; the boundary threshold (usually set to 0.2) forces the similarity of the positive samples to be at least 0.2 higher than that of the negative samples to avoid the model confusing easily confused categories; the cosine similarity is used to quantify the similarity of two semantic vectors.

[0108] Specifically, through contrastive learning, the model can better distinguish easily confused legal issues such as "patent infringement" and "trademark infringement", ensuring that the recommended clauses are highly relevant to the user's needs.

[0109] Specifically, contrastive learning forces the distinction between positive and negative samples to ensure that the recommended clauses are highly relevant to the user's question. For example, it forces the similarity of the positive samples (relevant clauses) to be higher than that of the negative samples (irrelevant clauses) by at least 0.2 (such as 0.8 vs 0.5) to avoid confusion; the cosine similarity quantifies semantic relevance and replaces traditional keyword matching. Using contrastive learning and cosine similarity ensures that the recommended content highly matches the user's needs.

[0110] The infringement probability generation module is used to construct a knowledge graph of key clauses and perform multi-hop reasoning on the knowledge graph based on a graph neural network to obtain the infringement probability score of the key clauses.

[0111] In the embodiment of the present invention, when the infringement probability generation module executes the construction of the knowledge graph of key clauses, it includes:

[0112] Extract the infringement elements of historical cases from the China Judgments Online.

[0113] Determine the nodes of the knowledge graph according to the infringement elements, historical cases and key clauses.

[0114] Mark the legal clause citation relationship and identify the association strength of the infringement elements.

[0115] Determine the edges of the knowledge graph according to the citation relationship and the association strength, thereby generating the knowledge graph of key clauses.

[0116] Specifically, use Stanford CoreNLP to automatically mark the legal clause citation relationship.

[0117] Specifically, the nodes of the knowledge graph represent cases, legal clauses and infringement elements (degree of technical feature overlap, trademark similarity); the edges of the knowledge graph represent the case-clause citation relationship and the element association strength.

[0118] Furthermore, the knowledge graph integrates legal provisions, historical cases, and infringement elements (such as "degree of overlap of technical features"); marks citation relationships (such as case A citing provision B) and association strength (a co-occurrence frequency > 5 times is a strong association), providing structured data support for the graph neural network and enhancing the credibility of infringement probability reasoning.

[0119] Specifically, the multi-hop reasoning aggregates the semantic information of nodes within 3 hops through the neighborhood message passing mechanism of the graph neural network to calculate the infringement probability score of the provision node.

[0120] Specifically, legal provision nodes are embedded with BERT vectors of the provision text, and historical case nodes are embedded with case descriptions; the edges include "provision A is cited by case B" and "cases C and D involve the same type of infringement".

[0121] In the embodiment of the present invention, the calculation formula of the infringement probability score in the infringement probability generation module is as follows:

[0122]

[0123] where P 侵权 is the infringement probability score, σ(*) is the Sigmoid function, W a is the weight for adjusting the output of the graph neural network, is to aggregate the current node feature and its neighbor node features, h v is the current node feature, is the neighbor node feature, b is the bias term in the graph neural network, v is the identifier of the current node, is the set of neighbor nodes of the current node v.

[0124] Specifically, the infringement probability score represents the likelihood that a certain provision is cited in an infringement case, 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 of a legal provision or case (such as 256-dimensional); the set of neighbor node features includes directly related cases, infringement elements, etc.; the bias term is used to enhance the flexibility of the model.

[0125] Specifically, through the multi-hop reasoning of the graph neural network (such as analyzing the association between "technical feature overlap" and multiple cases), the infringement risk of a certain provision can be predicted, providing a quantitative basis for legal advice.

[0126] In the embodiment of the present invention, the aggregation function adopted by the graph neural network in the infringement probability generation module is:

[0127]

[0128] where is the feature of the current node v at the k-th layer, ReLU(*) is the rectified linear unit activation function, W b is the weight matrix for fusing the features of the current node and its neighbor nodes, CONCAT(*) is the vector concatenation operation, is the feature of the current node v at the (k - 1)-th layer, MEAN(*) is the mean operation, is the feature of the neighbor node u at the (k - 1)-th layer, u is the identifier of the neighbor node, is the set of neighbor nodes of the current node v, k is the current layer number of the graph neural network, and v is the identifier of the current node.

[0129] Specifically, the information of neighboring nodes is aggregated through a graph neural network (GNN) to calculate the infringement probability; the Sigmoid function maps the output of the GNN to a probability value (e.g., 0.75 indicates a high risk); the mean operation fuses the features of neighbor nodes (such as the losing rate of related precedents), enhances the non-linear expression ability through ReLU activation, quantifies the infringement risk of clauses, and supports multi-dimensional legal suggestions (such as "high-risk clauses need to be avoided with emphasis").

[0130] Specifically, is the output of the previous layer; the mean operation refers to taking the average of the features of neighbor nodes to reduce the influence of noise; k is usually set to 2 layers to balance the model depth and the overfitting risk; the neighbor set of node v can be all precedents citing the same legal clause.

[0131] Specifically, by aggregating the features of neighbor nodes (such as the losing rate in precedents and the result of technical feature comparison), the contextual relevance of legal clauses can be captured, and the accuracy of infringement probability prediction can be improved.

[0132] Specifically, the knowledge graph integrates legal clauses, precedents, and infringement elements to construct a multi-dimensional association network; the GNN aggregates neighbor node information (such as the losing rate and technical overlap degree) and outputs an infringement probability score.

[0133] The clause screening module is used to mark high-risk clauses in the key clause set based on the infringement probability score, and perform clause screening on the key clause set in combination with semantic relevance to obtain the screened clauses.

[0134] Specifically, if a clause is cited in historical precedents and the losing rate is greater than 70%, it is additionally marked as "high risk".

[0135] 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 screened clauses, it includes: screening the final clauses in combination with the infringement probability and semantic relevance, where 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 excluded.

[0136] The reply generation module is used to generate a structured standard reply to the intellectual property consultation question according to the legal effect level of the screened clauses.

[0137] In an embodiment of the present invention, the conflict resolution rule for the legal effect level in the reply generation module is:

[0138] Sort the screened clauses according to the legal effect level, and the conflict content of the lower-level clauses is automatically covered by the higher-level clauses, where the legal effect level is: law prevails over regulations, and regulations prevail over precedents;

[0139] When the judicial interpretation conflicts with the precedent, the latest revised version of the judicial interpretation is adopted.

[0140] Specifically, sort according to the legal effect level (law > regulations > precedents) to ensure the authority of the advice, and the higher level automatically covers the conflict content of the lower level; when the judicial interpretation conflicts with the precedent, the latest revised version of the judicial interpretation is adopted to solve the clause conflict problem, avoid self-contradictory legal advice, and enhance user trust.

[0141] Refer to Figure 2 As shown, it is a schematic flowchart 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:

[0142] S1. Obtain the intellectual property consultation question of the user and associate the legal texts in the legal text database;

[0143] S2. Extract the context features of the legal text according to the multi-scale text convolution kernel;

[0144] S3. Focus and fuse the context features through the multi-head attention mechanism to obtain the key clause set of the legal text;

[0145] S4. Calculate the semantic relevance between the intellectual property consultation question and the key clauses based on the contrast learning algorithm;

[0146] S5. Construct a knowledge graph of the key clauses, and perform multi-hop reasoning on the knowledge graph based on the graph neural network to obtain the infringement probability score of the key clauses;

[0147] S6. High-risk clauses marked in the key clause set based on the infringement probability score are used, and clause screening is performed on the key clause set in combination with semantic relevance to obtain screened clauses;

[0148] S7. Structured standard responses to intellectual property consultation questions are generated according to the legal effect levels of the screened clauses.

[0149] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and the present invention can be implemented in other specific forms without departing from the spirit or basic characteristics of the present invention.

[0150] The embodiments of the present application can acquire and process relevant data based on artificial intelligence technology. Among them, artificial intelligence is a theory, method, technology, and application system that uses a digital computer or a machine controlled by a digital computer to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use knowledge to obtain the best results.

[0151] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions 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 relevance calculation module, an infringement probability generation module, a clause screening module, and a response generation module, where: The data acquisition module is used to obtain the user's intellectual property consultation questions and associate with the legal texts in the legal text database; The feature extraction module is used to extract the context features of legal texts according to multi-scale text convolution kernels. The construction method of the multi-scale text convolution kernels includes: setting 3 kernel sizes, which are respectively used to capture short clauses, paragraphs and cross-chapter semantics of legal texts; adopting a gating mechanism to suppress noise features, where the gating function of the gating mechanism is: Output = σ(W g ·x + b g ). In the formula, Output is the context feature tensor, x is the word vector of the legal text, b g is the bias term of the multi-scale text convolution, Conv(x) is the convolution operation, W g is the weight matrix for adjusting the importance of different semantic features, and σ(*) is the Sigmoid function; The feature fusion module is used to focus and fuse the context features through the 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 consultation question and the key clauses based on the contrast learning algorithm; The infringement probability generation module is used to construct a knowledge graph of the key clauses and perform multi-hop reasoning on the knowledge graph based on the graph neural network to obtain the infringement probability score of the key clauses; The clause screening module is used to mark the high-risk clauses in the set of key clauses based on the infringement probability score, and combine the semantic relevance to screen the clauses in the set of key clauses to obtain the screened clauses; The response generation module is used to generate a structured standard response to the intellectual property consultation question according to the legal effect level of the screened clauses.

2. The intelligent consultation response system based on deep learning according to claim 1, wherein When the data acquisition module executes to obtain the user's intellectual property consultation question and associate with the legal texts in the legal text database, it includes: Extract the legal entities in the intellectual property consultation question; Match the legal entities with the legal database to obtain the legal text corresponding to the intellectual property consultation question.

3. The intelligent consultation response system based on deep learning according to claim 1, wherein When the feature fusion module executes to focus and fuse the context features through the multi-head attention mechanism to obtain a set of key clauses of the legal text, it includes: Generate a query matrix, a key matrix, and a value matrix by linearly transforming the context features; Based on the query matrix, the key matrix, and the value matrix, calculate the attention weights of the attention heads one by one; Based on the attention weights, splice the output features of the attention heads to generate a fused feature matrix; Reduce the dimension of the fused feature matrix through a fully connected layer and output the key clause feature vectors; Based on the L2 norm sorting of the key clause feature vectors, retain the top N vectors as the set of key clauses of the legal text.

4. The intelligent consultation response system based on deep learning according to claim 1, wherein The contrast learning algorithm in the relevance calculation module is: L = max(0, sim(q, p - ) - sim(q, p + ) + γ) Where L is the semantic relevance loss, q is the embedded vector of the intellectual property consultation question, p - is the negative sample, γ is the margin threshold, p + is the positive sample, and sim(*) is the cosine similarity.

5. The intelligent consultation response system based on deep learning according to claim 1, characterized in that, When the infringement probability generation module executes to construct the knowledge graph of the key clauses, it includes: Extract the infringement elements of historical cases from the China Judgments Online; Determine the nodes of the knowledge graph according to the infringement elements, historical cases, and key clauses; Mark the legal clause citation relationship and identify the association strength of the infringement elements; Determine the edges of the knowledge graph according to the citation relationship and the association strength, thereby generating the knowledge graph of the key clauses.

6. The intelligent consultation response system based on deep learning according to claim 1, characterized in that, The calculation formula of the infringement probability score in the infringement probability generation module is as follows: Wherein, P 侵权 is the infringement probability score, σ(*) is the Sigmoid function, W a is the weight for adjusting the output of the graph neural network, is to aggregate the current node feature and its neighbor node features, h v is the current node feature, is the neighbor node feature, b is the bias term in the graph neural network, v is the identifier of the current node, is the set of neighbor nodes of the current node v.

7. The intelligent consultation response system based on deep learning according to claim 1, characterized in that, The aggregation function adopted by the graph neural network in the infringement probability generation module is: Among them, is the feature of the current node v at the k-th layer, ReLU(*) is the rectified linear unit activation function, W b is the weight matrix for fusing the features of the current node and its neighbor nodes, CONCAT(*) is the vector concatenation operation, is the feature of the current node v at the (k - 1)-th layer, MEAN(*) is the mean operation, is the feature of the neighbor node u at the (k - 1)-th layer, u is the identifier of the neighbor node, is the set of neighbor nodes of the current node v, k is the current layer number of the graph neural network, and v is the identifier of the current node.

8. The intelligent consultation response system based on deep learning according to any one of claims 1 to 7, characterized in that, The conflict resolution rule of the legal effect level in the response generation module is: Sort the screened clauses according to the legal effect level, and the high-level clauses automatically cover the conflicting content of the low-level clauses, where the legal effect level is: law is superior to regulations, and regulations are superior to cases; When the judicial interpretation conflicts with the case, the latest revised version of the judicial interpretation is adopted.

9. An intelligent consultation response method based on deep learning, characterized in that, The method includes: S1. Obtain the intellectual property consultation questions of the user and associate the legal texts in the legal text database; S2. Extract the context features of legal texts according to the multi-scale text convolution kernel. The construction method of the multi-scale text convolution kernel includes: setting 3 kernel sizes, which are respectively used to capture short clauses, paragraphs and cross-chapter semantics of legal texts; adopting a gating mechanism to suppress noise features. The gating function of the gating mechanism is: Output = σ(W g ·x + b g ). ·Conv(x), where Output is the context feature tensor, x is the word vector of the legal text, b g is the bias term of the multi-scale text convolution, Conv(x) is the convolution operation, W g is the weight matrix that adjusts the importance of different semantic features, and σ(*) is the Sigmoid function; S3. Focus and fuse the context features through the multi-head attention mechanism to obtain the set of key clauses of the legal text; S4. Calculate the semantic relevance between the intellectual property consultation question and the key clauses based on the contrastive learning algorithm; S5. Construct a knowledge graph of the key clauses and perform multi-hop reasoning on the knowledge graph based on the graph neural network to obtain the infringement probability score of the key clauses; S6. Mark the high-risk clauses in the set of key clauses based on the infringement probability score, and combine the semantic relevance to screen the clauses in the set of key clauses to obtain the screened clauses; S7. Generate a structured standard response to the intellectual property consultation question according to the legal effect level of the screened clauses.

Images