Patent agency whole-process intelligent management method based on artificial intelligence and big data

Abstract

The invention provides a patent agency whole-process intelligent management method based on artificial intelligence and big data, and the method comprises the steps: carrying out the real-time access of a global patent database, a legal status platform and a customer document through the collection of multi-source heterogeneous data and the construction of a dynamic patent knowledge graph, and storing technical feature entities, legal clause nodes and the like through a Neo4j graph database, calculating an edge weight in combination with the semantic vector to realize knowledge association; text feature dimension reduction and technical scheme similarity calculation are carried out through an improved TF-IDF algorithm and an LSTM-Attention model, and context dependency is effectively captured; and optimizing a reply strategy based on a deep reinforcement learning framework (DQN), defining an opinion feature vector and an action space, and dynamically adjusting a reply decision through a reward function (delta P is an authorization probability improvement amplitude, and T is time cost). The method is suitable for patent retrieval, technical scheme analysis, reply strategy generation and process automatic management.

Description

Technical Field

[0001] The present invention belongs to the technical field of patent information processing and intelligent decision-making, and specifically relates to a full-process intelligent management system and method for patent agency that integrates artificial intelligence algorithms and big data analysis, and is applicable to patent retrieval, technical solution analysis, response strategy generation, and process automation management. Background Art

[0002] With the rapid increase in the global patent application volume, the traditional patent agency process relies on manual experience to process multi-source heterogeneous data (such as patent documents, legal status texts, customer requirement documents, etc.), and there are problems such as low efficiency, poor consistency, and long response cycles. In the prior art, patent retrieval is mostly based on keyword matching (such as the Boolean retrieval model), and it is difficult to accurately capture the semantic associations of technical solutions (Smith J, Doe R. Patent analysis using keyword expansion [J]. Journal of Intellectual Property, 2020, 15(3): 45-60). The calculation of technical similarity often uses shallow algorithms such as cosine similarity, ignoring the context dependence of technical features (Zhang L et al. Semantic patent retrieval based on entity recognition [J]. IP Management Review, 2019, 12(2): 78-92).

[0003] Although existing automated systems (such as PatentOptimizer, PatSnap) can assist in process management, they lack the ability to dynamically optimize decisions on opinions. In addition, the construction of patent knowledge graphs mostly uses static relationship modeling and does not combine real-time legal status updates and semantic evolution characteristics (Chen H et al. Dynamic patent knowledge graph construction [J]. AI in Law, 2021, 28(4): 112-130). Therefore, there is an urgent need for a full-process management system that integrates multi-modal data processing, intelligent decision-making, and dynamic optimization. Summary of the Invention

[0004] The purpose of the present invention is to provide a full-process intelligent management system and method for patent agency based on artificial intelligence and big data, and through multi-source data fusion, in-depth feature mining, and dynamic decision optimization, to achieve full-process automation management of patent applications and responses. The specific technical solutions include the following steps:

[0005] Step 1: Multi-source heterogeneous data collection and patent knowledge graph construction

[0006] Step 1.1, access the global patent databases (WIPO, USPTO), legal status update platform, and customer requirement documents in real time through the API interface, and parse structured and unstructured data;

[0007] Step 1.2, construct a patent knowledge graph, define technical feature entities and legal clause nodes, and store them in the Neo4j graph database. The edge weight calculation is as follows:

[0008] w(e ij ) = α·Co-occur(i,j) + β·cos(v i ,v j )

[0009] where α = 0.6 and β = 0.4 are weight coefficients, and v i is the semantic vector of entity i.

[0010] Step 2: Data preprocessing and feature dimensionality reduction

[0011] Step 2.1, use the improved TF-IDF algorithm to weight the features of the patent text. The calculation formula is:

[0012]

[0013] where ∈ = 0.00001 is the smoothing coefficient, α = 0.5 is the adjustment factor, and θ t is the deflection angle of term t in the word vector space;

[0014] Step 2.2, generate a structured feature matrix, with the dimension reduced from the original 2000 dimensions to 300 dimensions.

[0015] Step 3: Intelligent calculation of technical solution similarity

[0016] Step 3.1, construct an LSTM neural network model, and the unit structure is defined as follows:

[0017] Input gate: i t = σ(W i ·[h t-1 ,x t +b i )

[0018] Forget gate: f t = σ(W f ·[h t-1 ,x t +b f )

[0019] Output gate: o t = σ(W o ·[h t-1 ,x t+b o )

[0020] Candidate memory unit:

[0021] Step 3.2, calculate the context weight by integrating the Attention mechanism:

[0022]

[0023] α is the alignment model function, in the form of bilinear transformation: a(s, h) = s T W a h, W a are trainable parameter matrices;

[0024] T is the sequence length;

[0025] e ij : At the i-th step of the decoder, the alignment score of the hidden state at the j-th position of the encoder, reflecting the correlation between the two;

[0026] s i-1 : The hidden state (vector) of the decoder at the (i - 1)-th step, as the "query", with dimension d s ;

[0027] h j : The hidden state (vector) of the encoder at the j-th step, as the "key", with dimension d h ;

[0028] α ij : At the i-th step of the decoder, the attention weight of the j-th position of the encoder, indicating the attention intensity to the j-th input when generating the i-th output;

[0029] exp(e ij ): Exponentiate the alignment score to ensure the result is positive;

[0030] Sum the exponential alignment scores for all time steps (k = 1 to T) of the encoder to achieve normalization;

[0031] c i : The context vector of the decoder at the i-th step, with the same dimension as h j same (d h );

[0032] Perform a weighted sum of the hidden states of the encoder for all time steps.

[0033] Step 4: Dynamically optimize the response strategy

[0034] Step 4.1: Using the deep reinforcement learning framework (DQN), define the state s as the opinion feature vector, and the actions a include three types of strategies: claim modification, evidence submission, and appeal.

[0035] Step 4.2: Design the Q-value update function:

[0036] Q(s,a)←Q(s,a)+η[r+γmax a′ Q(s′,a′)-Q(s,a)]

[0037] where η = 0.001 is the learning rate, γ = 0.95 is the discount factor, and the reward function r

[0038] is calculated by the following formula:

[0039] r=λ·ΔP-(1-λ)·T

[0040] where ΔP is the increase in the authorization probability, T is the time cost, and λ = 0.7 is the balance coefficient.

[0041] The further optimization solutions include:

[0042] The patent knowledge graph dynamically synchronizes legal status data every 24 hours;

[0043] The dimension of the hidden layer of the LSTM network is set to 256, the Dropout rate is 0.3, and the number of training rounds is 50;

[0044] The DQN network structure is a fully connected layer [512, 256], and the ReLU activation function is used.

[0045] Compared with the prior art, the beneficial effects of the present invention are:

[0046] Multi-source data fusion and dynamic knowledge graph: Construct a patent knowledge graph that is updated in real time, integrating global patent databases, legal status data, customer needs, and industry dynamics. Store technical features and legal clauses through the Neo4j graph database storage technology. The edge weights are combined with the co-occurrence frequency and semantic similarity calculation to break through the limitations of traditional static knowledge graphs and improve data relevance and retrieval accuracy.

[0047] Improved text feature extraction and dimensionality reduction: Adopt an improved TF-IDF algorithm that fuses word vector semantic information, and optimize the weight calculation by adjusting the factor α and the term deflection angle θ t to effectively capture the semantic differences of different types of texts, reduce the original 2000-dimensional features to 300 dimensions, and improve the efficiency of subsequent analysis.

[0048] The LSTM-Attention model enhances context understanding: By combining the Long Short-Term Memory network (LSTM) with the attention mechanism, it breaks through the limitations of shallow algorithms such as traditional cosine similarity, can capture the context dependence of technical solutions, accurately calculate the similarity between the proposed solution and the prior art, and reduce the risk of misjudgment (e.g., a similarity score of 0.92 in the embodiment triggers a high-risk warning).

[0049] Deep reinforcement learning optimizes the response strategy: Based on the deep Q-network framework, it dynamically optimizes the response decision path, balances the increase in the authorization probability and the time cost through the reward function, realizes the generation of intelligent strategies, and significantly improves the response efficiency and decision-making quality.

[0050] The closed-loop feedback mechanism enables continuous optimization: Collects actual patent application result data, feeds it back to the data preprocessing, intelligent analysis, and process management modules, and dynamically adjusts the model parameters to ensure that the system adapts to the changing legal environment and technological evolution.

[0051] Full-process automation and efficient resource utilization: Achieves full-process automation from data collection, retrieval and analysis to response strategy generation, reduces manual intervention, and improves processing consistency. Experimental data shows that the system occupies less than 8GB of video memory on an NVIDIA V100 GPU, and the response time is less than 3 seconds per case, combining high efficiency and low resource consumption. Brief Description of the Drawings

[0052] Figure 1 The overall system architecture diagram shows the interaction process of the data collection, preprocessing, analysis, and process management modules.

[0053] Figure 2 The structural diagram of the technical solution similarity calculation model includes the combination method of the LSTM and Attention mechanisms.

[0054] Figure 3 The state-action mapping schematic diagram of the deep Q-network in the response decision-making. Detailed Description of the Invention

[0055] The technical solution of the present invention will be further described below in conjunction with the embodiments. Those not specified in the embodiments are carried out according to conventional conditions, and the devices not specified in the manufacturer are regarded as conventional devices that can be obtained through the market.

[0056] Embodiment 1: Construction of Patent Knowledge Graph and Data Collection

[0057] Step 1.1: Multi-source data access and parsing

[0058] Data source configuration: Access the WIPO Global Patent Database (API 3.0), CNIPA, and customer requirement documents (PDF / Word format);

[0059] Data parsing: The BERT-BiLSTM-CRF model extraction technology is used to extract technical feature entities (such as "quantum computing chip", "blockchain consensus algorithm") and parse the claim validity tags in the legal status text;

[0060] Entity mapping: Define the association between the technical feature entity nodes and the legal clause nodes of Article 22 (creativity clause) and Article 26 (specification support clause) of the Patent Law.

[0061] Step 1.2: Knowledge graph storage and weight calculation

[0062] Database configuration: Use the Neo4j graph database, and the node types include technical features (more than 5,000) and legal clauses (more than 120), etc.;

[0063] Edge weight calculation:

[0064] w(e ij ) = 0.6·Co-occur(i,j) + 0.4·cos(v i ,v j )

[0065] where v i is a 300-dimensional semantic vector trained based on Word2Vec.

[0066] Example 2: Calculation of technical solution similarity

[0067] Step 2.1: Text feature extraction and dimensionality reduction

[0068] Input data: The solution to be applied, "Medical data privacy protection system based on federated learning", and the comparative document US2022000000A1;

[0069] Feature weighting: Use the improved TF-IDF algorithm, set ∈ = 0.00001, α = 0.5, and generate a 300-dimensional feature vector;

[0070] Matrix output: The feature matrix is normalized to [-1,1] and input into the LSTM-Attention model.

[0071] Step 2.2: Calculation of the LSTM-Attention model

[0072] Network parameters:

[0073] The LSTM hidden layer has 256 units, the Dropout rate is 0.3, and the time step is 50;

[0074] Attention weight matrix Initialized to follow N(0,0.02);

[0075] Similarity calculation: The output score is 0.92 (threshold 0.8), triggering a "high - risk duplication" warning.

[0076] Example 3: Dynamic optimization of response strategies

[0077] Step 3.1: Initialization of the reinforcement learning environment

[0078] State definition: The opinion feature vector includes the number of cited documents (5), the number of claim amendments (3), and the citation frequency of legal articles (Article 22 cited 2 times);

[0079] Action space: {Modify the independent claim (a1), Supplement experimental data (a2), Lodge an appeal (a3)};

[0080] Reward function: r = 0.7·ΔP - (1 - 0.7)·T, ΔP = 15%, T = 7 days r = +8.4

[0081] Step 3.2: DQN decision - making and training

[0082] Network structure: Fully - connected layer [512, 256], ReLU activation, learning rate η = 0.001, discount factor γ = 0.95;

[0083] Policy selection: ∈ - greedy policy (∈ = 0.1), select action a2 (submit supplementary data), update Q - value:

[0084] Q(s,a2)←22.6 + 0.001[8.4 + 0.95max a′ Q(s′,a′)-22.6]

[0085] Example 4: Full - process automated testing

[0086] Step 4.1: System integration and verification

[0087] Test data: 100 invention patents (60 in the field of electronic information, 40 in the field of biomedicine);

[0088] Process configuration:

[0089] Data collection cycle: Real - time synchronization (legal status) / Daily batch processing (customer documents);

[0090] Similarity calculation response time: <3 seconds per piece;

[0091] DQN decision - making iteration times: 50 rounds per case.

[0092] Step 4.2: Performance metric output

[0093] Retrieval accuracy: The recall rate of the top-10 patents is 98.7% (82.3% for traditional methods);

[0094] Response efficiency: The average cycle is shortened to 4.2 months (7.1 months for the manual process);

[0095] Resource consumption: The GPU video memory occupancy is <8GB (NVIDIA V100).

[0096] Summary of the parameters of the preferred embodiment

[0097]

[0098] In the description of this specification, the selection of the embodiments is only used for the description of specific technical features. Those skilled in the art can adjust the parameter configuration or algorithm combination according to actual needs, and such adjustments still fall within the protection scope of the claims of the present invention.

[0099] Finally, it should be noted that the above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A patent agency full-process intelligent management method based on artificial intelligence and big data, characterized in that: The following steps are involved: Step 1: Establish a multi-source heterogeneous data collection module: obtain patent document data, legal status data and customer demand text data through the API interface; Use crawler technology to obtain industry dynamics data on relevant patents and competitor patent layout data from public websites; The patent knowledge graph is constructed using the Neo4j graph database. The node types of the graph include technical features and legal terms. The edge relationship weight is determined by the co-occurrence frequency and semantic similarity. The data obtained through the API interface and the data obtained through crawler technology are supplemented into the patent knowledge graph. Step 2, data preprocessing module: use the improved TF-IDF algorithm to perform feature extraction and dimensionality reduction on text data to generate a structured feature matrix; Step 3, intelligent analysis module: build a technical solution similarity calculation model that integrates LSTM neural network and Attention mechanism to calculate the similarity index between the proposed solution and the existing technology; Step 4, process management module: Generate patent application strategy based on similarity index and use reinforcement learning algorithm to dynamically optimize the response decision path; The reinforcement learning algorithm uses a deep Q network framework: Q(s,a)←Q(s,a)+η[r+γmax a′ Q(s′,a′)-Q(s,a)] The state s is defined as the current opinion feature vector, the action space a contains the set of reply strategies, the reward function r is determined by the increase in the authorization probability and the time cost, η is the learning rate, and γ is the discount factor; The deep Q network uses an experience replay mechanism to store historical data of the agent's interaction with the environment in an experience replay buffer, and randomly samples data from the buffer for learning during training; Step 5, feedback optimization module: collect the actual result data of the patent application and feed it back to the data preprocessing module, intelligent analysis module and process management module to readjust the parameters α and ε of the improved TF-IDF algorithm, the weight matrix W and bias term b of the LSTM neural network, and the learning rate η and discount factor γ of the reinforcement learning algorithm.

2. The management method according to claim 1, characterized in that: The term weight calculation formula of the improved TF-IDF algorithm in step 2 is The improved TF-IDF algorithm is as follows: The term weight calculation formula is defined as W(t,d)=[(tf(t,d)×log(N / df(t)+ε))]×(1+α×cos(θ t )) Where tf(t,d) represents the frequency of term t in document d, N is the total number of documents, df(t) is the number of documents containing term t, ε is the smoothing coefficient, α is the adjustment factor, and θ t It represents the deflection angle of term t in the word vector space, and sets different α and ε values ​​for different types of text data (patent document data, legal status data, customer demand text data); The derivation process is as follows: The term weight formula in the traditional TF-IDF algorithm is W tf-idf (t,d)=tf(t,d)×idf(t) In order to avoid calculation errors caused by df(t) being 0, a smoothing coefficient ε is introduced. Considering the semantic information of the term in the word vector space, the adjustment factor α and the deflection angle θ of the term t in the word vector space are introduced t , through 1+α×cos(θ t ) adjusts the traditional TF-IDF weight to obtain an improved term weight calculation formula.

3. The management method according to claim 1, characterized in that: The derivation process of the structural formula of the LSTM neural network in step 3 is as follows: The core of the LSTM neural network is the memory unit c t In order to control the information inflow, outflow and forgetting of the memory unit, the input gate i is introduced t 、Forget gate t and output gate o t ; Input Gate i t Used to determine the current input x t How much information should be updated to memory cell c? t In the example, the sigmoid function is used to convert the input x t and the hidden state h at the previous moment t-1 Mapping to the interval [0,1] i t =σ(W i ·[h t-1 ,x t ]+b i ) Where W i is the weight matrix of the input gate, b i is the bias term; Forget Gate t The memory cell c used to determine the last moment t-1 How much information should be forgotten in the , and the sigmoid function is also used, that is f t =σ(W f ·[h t-1 ,x t ]+b f ) Where W f is the weight matrix of the forget gate, b f is the bias term; Candidate memory cells According to the current input x t and the hidden state h at the previous moment t-1 The calculated value is mapped to the interval [-1,1] using the hyperbolic tangent function, that is, Where W c is the weight matrix of the candidate memory unit, b c is the bias term; Memory unit update formula Indicates that, through the forget gate f t Decide to keep the memory unit c at the previous moment t-1 Part of the information in is passed through the input gate i t Deciding to add candidate memory cells Part of the information in, thereby updating the current memory unit c t ; Output gate o t Used to determine the current memory cell c t How much information should be output to the current hidden state h t In the above example, the sigmoid function is used, that is, the t =σ(W o ·[h t-1 ,x t ]+b o ) Where W o is the weight matrix of the output gate, b o is the bias term; Current hidden state h t =o t ⊙tanh(c t ) indicates that through the output gate o t Control memory unit c t The output of the information in , where ⊙ represents element-by-element multiplication; σ is the sigmoid activation function, and tanh is the hyperbolic tangent function; And the LSTM neural network uses batch normalization technology during training to normalize the features input to each layer. The formula is where x i is the input feature, μ is the mean, σ 2 is the variance, ∈ is a small constant to prevent division by zero; The calculation formula of the Attention mechanism is: e ij =a(s i-1 ,h j ) in α is the alignment model function, which adopts the bilinear transformation form: a(s,h)=s T W a h,W a is the trainable parameter matrix; T is the sequence length; e ij : The alignment score of the decoder to the hidden state of the encoder at the jth position at the i-th step, reflecting the correlation between the two; s i-1 : The hidden state (vector) of the decoder at step i-1, as the "query", with dimension d s ; h j : The hidden state (vector) of the encoder at step j, as the "key", with dimension d h ; α ij : The attention weight of the jth position of the encoder at the i-th step of the decoder, indicating the intensity of attention to the j-th input when generating the i-th output; exp(e ij ): Perform an exponential operation on the alignment score to ensure that the result is a positive value; Normalize by summing the index alignment scores of all encoder time steps (k=1 to T); c i : The context vector of the decoder step i, with the same dimension as h j Same (d h ); Take a weighted sum of the hidden states of the encoder at all time steps.

4. The management method according to claim 1, characterized in that: The deep Q network framework Q(s,a)←Q(s,a)+η[r+γmax a′ The derivation of Q(s′,a′)-Q(s,a)] is based on the Bellman equation; In the Markov decision process, the optimal action value function Q for taking action a in state s is * (s,a) satisfies the Bellman optimal equation Where r is the immediate reward, γ is the discount factor, and s′ is the next state; Q * (s,a): The optimal action value function for taking action a in state s, which represents the expected long-term cumulative reward that can be obtained by taking the optimal strategy based on the current state; r: ImmediateReward, which represents the instantaneous reward value (scalar) of the environment feedback after executing action a; γ: Discount factor (DiscountFactor), value range [0,1), used to balance the importance of current rewards and future rewards (the closer γ is to 1, the more the model values ​​long-term benefits); s′: NextState, the state to which the action a is transferred after executing it; max a′ Q * (s′, a′): choose the state that maximizes Q in the next state s′ * The action a′; The expectation operator, which represents the average of the randomness of state transitions and rewards in the environment; In the deep Q network, a neural network Q(s′,a′;θ - ) to approximate the optimal action-value function Q * (s,a), update the network parameters θ by minimizing the loss function, where θ - are the parameters of the target network; The minimization loss function is: θ: The parameters of the current Q network (neural network weights), used to estimate Q(s,a); θ - : The parameters of the target Q network are copied from the current Q network periodically (parameter update lag) and are used to calculate the target value r+γmax a′ Q(s′,a′;θ - ); Q(s,a;θ): the current Q network’s estimate of the value of state s and action a (scalar); Q(s′,a′;θ-): the target Q-network’s estimate of the value of the next state s′ and action a′ (scalar); L(θ): Mean square error (MSE) loss function, which measures the gap between the current network prediction value and the target value; Use stochastic gradient descent method to update parameters and get parameter update formula in θ: the parameters of the current Q network (the neural network weights to be updated); η: learning rate (LearningRate), controls the parameter update step size (scalar); The gradient (directional derivative) of the loss function with respect to the parameters θ indicates how to adjust θ to reduce the loss.

Images