Patent process scheduling monitoring method and system, storage medium and program product

Abstract

The application discloses a patent process scheduling monitoring method, system, storage medium and program product, belongs to the artificial intelligence technical field, and the patent process scheduling monitoring method comprises the following steps: identifying a user question for a patent process to obtain a user intention; decomposing and arranging the user intention to obtain an arrangement task for the patent process and a work intelligent agent corresponding to the arrangement task; the arrangement task comprises a query task of the patent process and / or a monitoring task of the patent process; the work intelligent agent executes the arrangement task according to a pre-trained patent process large model to obtain a task result corresponding to the arrangement task; wherein the patent process large model is used for outputting a query result of the patent process according to the query task of the patent process and / or outputting a monitoring result of the patent process according to the monitoring task of the patent process. The application improves the processing execution efficiency by calling the patent process large model to query and / or monitor the patent process through the work intelligent agent.

Description

Technical Field

[0001] This application belongs to the field of artificial intelligence technology, specifically relating to a patent process scheduling and monitoring method, system, storage medium, and program product. Background Technology

[0002] Patent processes involve internal proposals, cross-entity reviews, dynamic data changes within the process, and collaborative approvals from external units. Compared to ordinary internal workflows, these processes are more complex and time-consuming. In process management systems, due to the large number of approval personnel and the complexity of branch processes and countersignatures, quickly locating process approval progress, dynamically analyzing process execution efficiency, and providing process risk warnings are key issues that current process management systems must address. Existing technologies lack the ability to track the entire process chain using large models and lack real-time monitoring and anomaly detection mechanisms for process instances and task status during patent process query and monitoring, resulting in low processing efficiency. Summary of the Invention

[0003] This application provides a method, system, storage medium, and program product for scheduling and monitoring patent processes, in order to solve the problem that existing patent process scheduling and monitoring methods cannot perform full-link tracking and query monitoring through large models.

[0004] In a first aspect, embodiments of this application provide a method for scheduling and monitoring a patent process, comprising:

[0005] Identify user questions related to the patent process to obtain user intent;

[0006] The user intent is decomposed and orchestrated to obtain orchestration tasks for the patent process and corresponding working agents for the orchestration tasks; wherein, the orchestration tasks include patent process query tasks and / or patent process monitoring tasks.

[0007] The working agent executes the orchestration task according to the pre-trained patent process big model to obtain the task result corresponding to the orchestration task; wherein, the patent process big model is used to output the query result of the patent process according to the query task of the patent process and / or output the supervision result of the patent process according to the monitoring task of the patent process.

[0008] Optionally, the method further includes:

[0009] Acquire training data for the patent process, wherein the training data includes: patent process template data, data of organizations that file patents, patent proposal data, and patent process chart data;

[0010] Training targets are generated based on the patent process query task and / or patent process monitoring task.

[0011] Based on the training data and the training objective, the initial model used for querying patent processes and / or for monitoring patent processes is trained to obtain the large patent process model.

[0012] Optionally, the initial model includes a transformation encoder and a recurrent neural network;

[0013] Specifically, based on the training data and the training objective, an initial model for querying and / or monitoring patent processes is trained to obtain the large-scale patent process model, including:

[0014] The patent flowchart data is processed using a conversion encoder to extract visual features, thereby obtaining image features.

[0015] The patent process template data is subjected to text feature processing to obtain the first text feature;

[0016] Text feature processing is performed on the organizational data of the patent-claiming organization to obtain a second text feature;

[0017] The patent proposal data is processed to obtain a third text feature;

[0018] Using the recurrent neural network, the initial model is trained based on the training objective, the image features, the first text features, the second text features, and the third text features to obtain the large model of the patent process.

[0019] Optionally, the step of using a transformation encoder to perform visual feature extraction processing on the patent flowchart data to obtain image features includes:

[0020] The patent flowchart data is segmented to obtain chart data blocks;

[0021] The chart data blocks are linearly mapped and flattened to obtain a mapping vector;

[0022] Add positional encoding to the mapping vector to obtain a sequence of chart data blocks;

[0023] The image features are obtained by processing the sequence of chart data blocks using the multi-head self-attention mechanism and feedforward neural network in the transformation encoder.

[0024] Optionally, the step of using the recurrent neural network to train the initial model based on the training objective, the image features, the first text features, the second text features, and the third text features to obtain the large-scale patent process model includes:

[0025] The image features, the first text features, the second text features, and the third text features are fused together to obtain fused features.

[0026] The recurrent neural network is used to obtain the temporal dependency of the fused features and the contextual information of the fused features;

[0027] By utilizing a bidirectional cross-attention mechanism, visual features for text perception and visually guided text features are obtained based on the temporal dependency of the fused features and the contextual information of the fused features.

[0028] A first scaling factor and a second scaling factor are obtained based on the similarity between the image features and the text features using a dynamic gradient gating mechanism; wherein the text features include the first text feature, the second text feature, and the third text feature;

[0029] The target text sequence is obtained based on the first scaling factor, the visual features of the text perception, the second scaling factor, and the visually guided text features;

[0030] Based on the loss information between the training objective and the target text sequence, the model parameters of the initial model are updated to obtain the large model of the patent process.

[0031] Optionally, after decomposing and orchestrating the user intent to obtain the orchestration task for the patent process and the working agent corresponding to the orchestration task, the method further includes:

[0032] The orchestration task is semantically decomposed to obtain at least one level of task statements;

[0033] The task statements at at least one level are semantically aligned according to preset semantic alignment rules to obtain the processed orchestration task; wherein, the preset semantic alignment rules include the association between task statements at different levels and the preset standard statements corresponding to each level of task statements.

[0034] Send the processed orchestration task to the working intelligent agent;

[0035] The step of executing the orchestration task by the working intelligent agent based on a pre-trained patent process large model to obtain the task result corresponding to the orchestration task includes:

[0036] The process involves the working agent executing the processed orchestration task based on a pre-trained patent process model to obtain the task result corresponding to the orchestration task.

[0037] Optionally, identifying user questions related to the patent process to obtain user intent includes:

[0038] The user's question undergoes a first processing step to obtain a question identification result; wherein, the first processing step includes natural language understanding processing, key sentence identification, and question entity identification;

[0039] The initial user intent is obtained by performing intent matching on the problem identification results.

[0040] The user intent is determined based on the initial user intent and the first information corresponding to the initial user intent;

[0041] The first information corresponding to the initial user intent includes the first confidence level of the initial user intent and the user's confirmation instruction information for the initial user intent.

[0042] Optionally, the step of executing the orchestration task by the working intelligent agent according to the pre-trained patent process large model to obtain the task result corresponding to the orchestration task includes:

[0043] The working intelligent agent sends task data to the patent process big model according to the orchestrated tasks;

[0044] The task result data output by the large model of the patent process is obtained through the working intelligent agent;

[0045] The task result is obtained by processing the task result data through the working intelligent agent.

[0046] Optionally, the working intelligent agent includes a natural language parsing intelligent agent, a graph processing intelligent agent, a risk warning intelligent agent, and a verification intelligent agent;

[0047] The task result data includes chart data and text data;

[0048] The step of processing the task result data through the working intelligent agent to obtain the task result includes:

[0049] The text data is processed using the natural language parsing agent, and the chart data is processed using the chart processing agent to obtain the initial task results and risk information during the processing.

[0050] The risk warning intelligent agent is used to monitor the risk information and obtain monitoring results;

[0051] The verification agent is used to verify the initial task result to obtain the verification result;

[0052] The task result is obtained based on the monitoring result, the verification result, and the initial task result.

[0053] Optionally, the method further includes:

[0054] Based on the second confidence level of the working agent and the task result data corresponding to the working agent, the third confidence level of the task result is obtained;

[0055] The accuracy of the task result is determined based on the third confidence level.

[0056] Secondly, embodiments of this application also provide a scheduling and monitoring system for a patent process, comprising:

[0057] A dialogue agent, which is used to identify user questions regarding the patent process and obtain user intent;

[0058] A coordinating and supervising intelligent agent is used to decompose and orchestrate the user intent to obtain an orchestration task for the patent process and a working intelligent agent corresponding to the orchestration task; wherein, the orchestration task includes a patent process query task and / or a patent process monitoring task.

[0059] A working intelligent agent is used to execute the orchestration task according to a pre-trained patent process big model to obtain the task result corresponding to the orchestration task; wherein, the patent process big model is used to output the query result of the patent process according to the query task of the patent process and / or output the supervision result of the patent process according to the monitoring task of the patent process.

[0060] Thirdly, embodiments of this application also provide a patent process scheduling and monitoring device, comprising: a processor, a memory, and a program stored in the memory and executable on the processor, wherein when the program is executed by the processor, it implements the steps of the patent process scheduling and monitoring method as described in any one of the first aspects.

[0061] Fourthly, embodiments of this application also provide a readable storage medium storing a program, which, when executed by a processor, implements the steps of the scheduling and monitoring method of the patent process as described in any one of the first aspects.

[0062] Fifthly, embodiments of this application also provide a computer program product, including computer instructions, which, when executed by a processor, implement the steps in the scheduling and monitoring method of the patent process as described in any one of the first aspects.

[0063] The beneficial effects of this application are:

[0064] The patent process scheduling and monitoring method provided in this application pre-trains a large-scale patent process model to output query results and / or monitoring results based on patent process query tasks. This enables end-to-end tracking of patent process queries and / or monitoring through the large-scale patent process model. The method identifies user questions related to the patent process to obtain user intent, decomposes and orchestrates this intent to generate orchestration tasks and corresponding working agents. These working agents execute the orchestration tasks according to the pre-trained large-scale patent process model, yielding the corresponding task results. By using the working agents to invoke the large-scale patent process model for querying and / or monitoring the patent process, processing efficiency is improved. Attached Figure Description

[0065] Figure 1 This is a flowchart of the scheduling and monitoring method for the patent process provided in the embodiments of this application;

[0066] Figure 2 This is one of the structural schematic diagrams of the scheduling and monitoring system for the patented process provided in this application embodiment;

[0067] Figure 3 This is a schematic diagram of the structure of the working agent provided in the embodiments of this application;

[0068] Figure 4 This is the second schematic diagram of the structure of the scheduling and monitoring system for the patent process provided in the embodiments of this application;

[0069] Figure 5 This is a schematic diagram of the structure of the scheduling and monitoring equipment for the patented process provided in the embodiments of this application. Detailed Implementation

[0070] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0071] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and are not used to describe a specified order or sequence. It should be understood that such use of data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first" and "second" are generally of the same class, not limited in number; for example, a first object can be one or more. Furthermore, in the specification and claims, "and" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0072] In the following description, specific details such as particular configurations and components are provided merely to aid in a comprehensive understanding of the embodiments of this application. Therefore, those skilled in the art will understand that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. Furthermore, for clarity and brevity, descriptions of known functions and constructions have been omitted.

[0073] It should be understood that the phrase "one embodiment" or "an embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "in one embodiment" or "in an embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments.

[0074] In the various embodiments of this application, it should be understood that the sequence number of each process described below does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0075] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and are not used to describe a specified order or sequence. It should be understood that such use of data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first" and "second" are generally of the same class, not limited in number; for example, a first object can be one or more. Furthermore, in the specification and claims, "and" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0076] Furthermore, the "or" in this application indicates at least one of the connected objects. For example, "A or B" covers three scenarios: Scenario 1: includes A but excludes B; Scenario 2: includes B but excludes A; Scenario 3: includes both A and B. The character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0077] The term "instruction" in this application can be either a direct instruction (or explicit instruction) or an indirect instruction (or implicit instruction). A direct instruction can be understood as one in which the sender explicitly informs the receiver of specific information, the operation to be performed, or the requested result, etc.; an indirect instruction can be understood as one in which the receiver determines the corresponding information based on the instruction sent by the sender, or makes a judgment and determines the operation to be performed or the requested result, etc., based on the judgment result.

[0078] Before describing the specific embodiments of this application, the following will be explained first:

[0079] In recent years, with the development of large language models, their superior instruction capabilities and contextual analysis and learning abilities have enabled them to complete various data processing tasks in numerous fields, significantly changing people's lives, learning, and work. Various applications based on large language models have greatly improved people's work efficiency.

[0080] To address the problem that existing patent process scheduling and monitoring methods cannot perform full-link tracking and query monitoring through large models, this application provides a patent process scheduling and monitoring method, system, storage medium, and program product.

[0081] like Figure 1 As shown in the figure, this application provides a method for scheduling and monitoring a patent process, including:

[0082] Step 101: Identify user questions related to the patent process to obtain user intent.

[0083] It should be noted that the patent process scheduling and monitoring method provided in this application embodiment is applied to a patent process scheduling and monitoring system, the structural diagram of which is shown below. Figure 2 As shown, the patent process scheduling and monitoring system includes a front-end interaction layer, which includes a dialogue agent. The patent process scheduling and monitoring system interacts with users through a dialogue interface. Users raise questions about the patent process through the dialogue interface, and the dialogue agent identifies the questions to obtain the user's intent (or intent result).

[0084] Step 102: Decompose and orchestrate the user intent to obtain orchestration tasks for the patent process and the corresponding working agents for the orchestration tasks; wherein, the orchestration tasks include patent process query tasks and / or patent process monitoring tasks.

[0085] In this embodiment, as Figure 2 As shown, the patent process scheduling and monitoring system also includes a coordination and supervision layer. This layer comprises a coordination and supervision agent, which in turn includes a coordinating agent and a supervision agent. The coordinating agent receives (or observes) user intents sent by the dialogue agent, decomposes and analyzes these intents, and orchestrates tasks to obtain query tasks and / or monitoring tasks for the patent process. The supervision agent parses the orchestration tasks and determines the working agents that need to be scheduled for each step of the orchestration task.

[0086] Specifically, the coordinating agent receives and decomposes tasks. After observing the user intent sent by the dialogue agent, it performs preliminary analysis and decomposition of the user intent to determine the main steps and objectives to be executed. At the same time, based on the main characteristics of the task, it orchestrates the task and obtains the agent information to be scheduled for each step. The task orchestration, agent information, user intent, and other parameters, as well as the agent information to be scheduled for each step, are used as input parameters and then passed to the supervising agent.

[0087] Additionally, the supervisory agent obtains parameters such as orchestration tasks and user intents sent by the coordinating agent, parses the orchestration tasks, and determines the working agents that need to be scheduled for each step of the orchestration task.

[0088] Step 103: The working agent executes the orchestration task according to the pre-trained patent process big model to obtain the task result corresponding to the orchestration task; wherein, the patent process big model is used to output the query result of the patent process according to the query task of the patent process and / or output the supervision result of the patent process according to the monitoring task of the patent process.

[0089] like Figure 2 As shown, the patent process scheduling and monitoring system also includes a data model layer, which stores a pre-trained large model of the patent process and training data of the patent process.

[0090] The data model layer, addressing the intelligent process management, intelligent process analysis management, and intelligent patent approval process risk warning issues to be addressed in this embodiment, adopts a unified dialogue mode as the system entry point. Communication with the user is accomplished through a dialogue agent, primarily recognizing the user's input intent and leveraging contextual information to establish the user's true intent. The results returned by the agent are then presented to the user.

[0091] In this step, a large-scale patent process model is first pre-trained based on the training data of the patent process. This model is used to output query results for patent process query tasks and / or supervision results for patent process monitoring tasks. This large-scale patent process model enables end-to-end tracking of patent process queries and / or monitoring.

[0092] The coordinating and supervising agent parses the orchestration task and sends scheduling instructions and orchestration tasks to the working agent. The scheduling instructions are used to instruct the working agent to call the patent process big model according to the orchestration task, execute the orchestration task, and obtain the task result corresponding to the orchestration task.

[0093] When there are two or more working agents, the coordinating and supervising agent determines the first working agent to be scheduled based on the orchestration task, and sends the orchestration task and user intent to the first working agent. The first working agent executes the orchestration task according to the patent process model. The task result of the first working agent's orchestration task is used as the input of the next working agent. The coordinating and supervising agent determines the next working agent to be scheduled based on the orchestration task, and sends the orchestration task, user intent, and the task result of the previous working agent to the next working agent. The above steps are repeated to perform multi-agent scheduling and switching. When the coordinating and supervising agent determines that there is no next step in the current orchestration task, it feeds back the task result to the coordinating and supervising agent, indicating that the current orchestration task has been fully executed.

[0094] This step involves using a working intelligent agent to call upon the large-scale patent process model to query and / or monitor the patent process, thereby improving processing efficiency.

[0095] In some embodiments of this application, the method further includes:

[0096] Obtain training data for the patent process.

[0097] Specifically, training data for the patent process is obtained through the data model layer, where the training data for the patent process is obtained from the patent process management system.

[0098] The training data includes:

[0099] The patent process template data is obtained from the full set of process templates extracted from the patent process management system. Specifically, it involves extracting the full set of process templates, including all process node information, connection information, configuration information, etc., such as extracting process instance data within the two years prior to the current time.

[0100] The data on the organization that filed the patent is specifically compiled and labeled with organizational data and personnel identity information to obtain the data on the organization that filed the patent. This data is then configured with process nodes to provide the required identification of process nodes for the large patent process model.

[0101] Specifically, patent proposal data is extracted from the patent process system using pre-defined anonymization standards. This data is then used to supplement the large-scale patent process model with identifiers for different types of proposals from different organizations.

[0102] Patent flowchart data, which is collected from the patent process management system, includes proposal charts, patent flowcharts, process structure diagrams, and process logic diagrams. This ensures the diversity and representativeness of the large patent process model and covers charts from different technical fields and patent types.

[0103] Training objectives are generated based on the patent process query task and / or patent process monitoring task.

[0104] Specifically, training objectives are defined based on the specific patent process. For example, in patent process queries, basic data such as the patent proposal organization, user data of the proposal, and proposal time are used as training objectives; and / or, in patent monitoring, normal patent data and abnormal patent numbers, overdue approvals, and patent process expiration dates are used as training objectives.

[0105] Based on the training data and the training objective, the initial model used for querying patent processes and / or for monitoring patent processes is trained to obtain the large patent process model.

[0106] Specifically, the training data is cleaned to remove noise and irrelevant information, and the patent flowchart data is labeled, including the type of flowchart data, semantic information of each component, and flowchart node information. Then, the patent flowchart data is standardized, such as adjusting image size and normalizing pixel values, to meet the model's input requirements. Simultaneously, the flowchart data is augmented, such as by flipping and scaling, to expand the dataset and improve the model's generalization ability. After the aforementioned processing of the training data, the initial model used for querying patent flows and / or monitoring patent flows is trained based on the processed training data and the training objective, resulting in the large-scale patent flow model.

[0107] Optionally, the initial model includes a transformation encoder and a recurrent neural network (RNN).

[0108] Specifically, based on the training data and the training objective, an initial model for querying and / or monitoring patent processes is trained to obtain the large-scale patent process model, including:

[0109] The patent flowchart data is processed using a conversion encoder to extract visual features, resulting in image features.

[0110] The transformation encoder is a Vision Transformer (ViT). That is, feature extraction is performed using the Vision Transformer (ViT).

[0111] Text feature processing is performed on the patent process template data to obtain a first text feature; text feature processing is performed on the data of the organization that filed the patent to obtain a second text feature; and text feature processing is performed on the patent proposal data to obtain a third text feature.

[0112] For example, in this embodiment, a pre-trained large model, such as the DeepSeek model, is used to perform text feature processing on the patent process template data, the data of the organization that filed the patent, and the patent proposal data to obtain the corresponding text features.

[0113] The initial model is trained using the recurrent neural network (RNN) based on the training objective, the image features, the first text features, the second text features, and the third text features to obtain the large model of the patent process.

[0114] Specifically, ViT, RNN, and the output layer are combined to form a complete initial training model. Using this RNN and output layer, the initial model is trained based on the training objective, image features, first text features, second text features, and third text features to obtain a trained large-scale patent process model.

[0115] In an optional implementation, the step of using a transformation encoder to perform visual feature extraction processing on the patent flowchart data to obtain image features includes:

[0116] The patent flowchart data is segmented to obtain chart data patches.

[0117] Specifically, the patent flowchart data is divided into multiple fixed-size chart data patches, such as H*W (where H represents the height of the chart data patch and W represents the width of the chart data patch), resulting in P*P smaller patches, each of which is a chart data patch.

[0118] The chart data blocks are linearly mapped and flattened to obtain a mapping vector.

[0119] Specifically, each chart data patch is flattened into a one-dimensional vector and mapped to the dimension required by the patent process large model through a linear layer, resulting in the vector representation corresponding to each chart data patch, i.e., the mapping vector.

[0120] Add positional encoding to the mapping vector to obtain a sequence of chart data blocks.

[0121] It should be noted that since the Transformer itself does not have the ability to process spatial information, it is necessary to add position encoding to the mapping vector of each chart data patch in order to preserve the spatial structure information of the image and obtain the chart data patch sequence.

[0122] The image features are obtained by processing the sequence of chart data blocks using the multi-head self-attention mechanism and feedforward neural network in the transformation encoder.

[0123] Specifically, the sequence of patches with added position encoding (i.e., the sequence of graph data blocks) is input into the Transformer encoder. The encoder processes the sequence through a multi-head self-attention mechanism and a feedforward neural network to capture the global associations and dependencies between different parts of the image and output the feature representation of each patch, i.e., the image features.

[0124] In an optional implementation, the step of using the recurrent neural network to train the initial model based on the training objective, the image features, the first text features, the second text features, and the third text features to obtain the large-scale patent process model includes:

[0125] The image features, the first text features, the second text features, and the third text features are fused together to obtain fused features.

[0126] Specifically, the chart image features extracted by ViT are fused with the text features (including the first text feature, the second text feature, and the third text feature) generated by the DeepSeek model. The fused feature vector (i.e., the fused feature) is obtained by concatenation, weighted summation, and other methods.

[0127] For example, image features, first text features, second text features, and third text features can be directly concatenated along the vector dimension to form fused features of varying lengths, as shown in the following formula:

[0128]

[0129] Where I represents the input patent flowchart data, T represents the text data (including patent flowchart template data, data on organizations that file patents, and patent proposal data), and ViT(·) and DeepSeek(·) output d_v-dimensional features (image features) and d_t-dimensional features (including first text features, second text features, and third text features), respectively. Norm(·) represents L2 normalization; The channel dimension is concatenated; α and β are learnable scalar weights that satisfy α+β=1 and 0≤α,β≤1, and are dynamically adjusted through backpropagation.

[0130] The recurrent neural network is used to obtain the temporal dependency of the fused features and the contextual information of the fused features.

[0131] Specifically, the fused features are input into an RNN for sequence modeling. RNNs can model sequential data, capturing temporal dependencies and contextual information within the sequence. In patent process handling, the fused feature sequence can be sequentially input into the RNN according to the order of the patent process. The hidden states of the RNN are continuously updated as time steps progress, thus forming a dynamic semantic understanding of the entire patent process.

[0132] By utilizing a bidirectional cross-attention mechanism, visual features for text perception and visually guided text features are obtained based on the temporal dependency of the fused features and the contextual information of the fused features.

[0133] Specifically, after fusing the temporal dependencies and contextual information of the fused features, bidirectional cross-attention is used. First, the sequence of chart data blocks is treated as a key / value pair and the text features as a query, and image-to-text attention is calculated to obtain text-aware visual features. Then, the text features are treated as a key / value pair and the sequence of chart data blocks as a query, and text-to-image attention is calculated to obtain visually guided text features. The two sets of outputs are weighted by Dynamic Gradient Gating (DGG) and fused with their respective original features. With the addition of a mutual information maximization alignment loss, accurate cross-modal alignment of images and text can be achieved by only updating the cross-attention parameters.

[0134] After feature fusion, during application, the fixed weights in traditional models may lead to modal imbalance, ultimately causing the large model to confuse the true intent. To address this issue, this application introduces dynamic gradient gating, which can adjust the gradient contribution ratio in real time based on the relevance of the image and text.

[0135] A first scaling factor and a second scaling factor are obtained based on the similarity between the image features and the text features using a dynamic gradient gating mechanism; wherein the text features include the first text feature, the second text feature, and the third text feature.

[0136] It should be noted that after feature fusion, during application, the fixed weights in traditional models may lead to modal imbalance, ultimately causing the large model to confuse the true intent. To address this issue, this application introduces dynamic gradient gating, which can adjust the gradient contribution ratio in real time based on the relevance of the image and text.

[0137] The image-text relevance vector associates image features and text features based on the preceding content, and then uses a cosine similarity algorithm to obtain their similarity value. For example, the formula is as follows:

[0138] s_t = (v_V·v_L) / (‖v_V‖‖v_L‖) ∈[-1,1]

[0139] Where v_V represents image features and v_L represents text features.

[0140] Based on the similarity value and the smoothed correlation function λ, the smoothed correlation coefficient ρ_t is obtained, as shown in the following formula:

[0141] ρ_t = λρ_{t-1}+(1-λ)s_t

[0142] Where λ=0.7 is determined based on empirical values, and t represents the number of training iterations of the patent process model. For each batch processed (the number of iterations increases by 1), t increases by 1 to achieve "gradual" modal equilibrium.

[0143] The dynamic gradient gating mechanism can learn a scaling factor based on the smoothed correlation coefficient ρ_t. The scaling factor includes a first scaling factor and a second scaling factor, and the specific formula is as follows:

[0144] ;

[0145] β_t = 1-α_t.

[0146] Where α_t represents the first scaling factor, β_t represents the second scaling factor, W_α represents the linear projection weight of the image-text relevance to the first scaling factor, used to control the influence of ρ_t on α_t, with an initial value of 0, indicating that the relevance difference is ignored in the early stage of training, b_α represents the bias term of the linear projection, which determines the initial gating position, initially set to 1, so that σ(1)≈0.73, and after subsequent conversion by β_t=1-α_t, α_t≈0.5. The set of real numbers that represents a 1-row × 1-column structure.

[0147] During the initialization phase of learning, W_α=0 and b_α=1, ensuring that α≈0.5 in the early stages of training, thus balancing the gradients of image features and text features. Here, α represents the gradient contribution weight.

[0148] The visual features of the text perception are dynamically weighted by a first scaling factor, and the visually guided text features are dynamically weighted by a second scaling factor.

[0149] Dynamic gradient gating is based on the "smoothing correlation coefficient ρ_t". It uses learnable α_t and β_t to scale the ViT and LLM gradients in real time during the backpropagation stage. This does not increase the forward delay (<0.3 ms) but also ensures that noise gradients are suppressed when the graph and text are weakly correlated and the interaction signal is amplified when they are strongly correlated, thus achieving true "on-demand contribution".

[0150] The target text sequence is obtained based on the first scaling factor, the visual features of the text perception, the second scaling factor, and the visually guided text features.

[0151] Specifically, ViT, RNN, and the corresponding output layer are combined into a complete training model. The output layer can be designed according to the training objective, using a fully connected layer with a softmax function to output the target text sequence. This target text sequence is obtained by dynamically weighting the visual features of the text perception using a first scaling factor and dynamically weighting the visually guided text features using a second scaling factor.

[0152] Based on the loss information between the training objective and the target text sequence, the model parameters of the initial model are updated to obtain the large model of the patent process.

[0153] The optimization algorithm uses the Adam algorithm, and the loss function is determined according to the training objective. The patented model mainly uses the cross-entropy loss function and the sequence-to-sequence loss function.

[0154] During model training, based on the processing tasks of the patent process, accuracy, precision, recall, F1 score, etc. are selected as evaluation indicators. The Bilingual Evaluation Understudy (BLEU) indicator is used to evaluate the quality of the generated text and the similarity to the reference text.

[0155] The trained patent process model was evaluated using a test set, and the values ​​of various evaluation metrics were calculated to understand the model's performance in real-world applications.

[0156] Based on the evaluation results of the large-scale patent process model, the model is optimized. This can be achieved by adjusting hyperparameters such as learning rate, batch size, and hidden layer dimension; improving the model's structure, such as increasing or decreasing the number of encoder layers in ViT or the number of hidden layer units in RNN; and employing regularization and early stopping techniques to prevent overfitting.

[0157] It should also be noted that during the scheduling and orchestration of agents, different agents may have different understandings of the nouns or terms input by the user, which may lead to significant deviations in the final results. In order to solve the problem of inconsistent understanding of the same term by different agents, this application proposes an ontology-based semantic alignment technology to realize the "semantic alignment" and "conflict resolution" mechanism between agents, thereby improving the consistency of results obtained by multiple agents.

[0158] In an optional embodiment, after decomposing and orchestrating the user intent to obtain an orchestration task for the patent process and a working agent corresponding to the orchestration task, the method further includes:

[0159] The orchestration task is semantically split to obtain at least one level of task statements.

[0160] For example, the orchestration task includes querying the patent status and determining whether the patent is valid or invalid, as well as the reason for invalidation. Semantic decomposition of the above orchestration task yields three levels of task statements: patent status, valid or invalid, and reason for invalidation.

[0161] The task statements at at least one level are semantically aligned according to preset semantic alignment rules to obtain the processed orchestration task; wherein, the preset semantic alignment rules include the association between task statements at different levels and the preset standard statements corresponding to each level of task statements.

[0162] Optionally, the preset semantic alignment rules can be stored in a graph database. For example, a preset graph database can be used to store concept levels, synonyms, attribute relationships, etc. The relationships in the graph database are as follows:

[0163] Patent status

[0164] |--- Valid

[0165] |--- Invalid

[0166] Unpaid annual fee

[0167] |--- Voluntary abandonment

[0168] The rejection took effect.

[0169] Using the aforementioned preset semantic alignment rules, semantic alignment processing is performed on the task statements at at least one level. When the agent generates and outputs orchestration tasks, it must pass the OntologyValidator module. The validation process includes:

[0170] Ensure that the terminology used by each agent for orchestration tasks conforms to the standard ontology area defined in the graph database above and conforms to the preset standard statements;

[0171] If a semantic conflict occurs (for example, the agent says "abort", but the default standard statement for the relation in the graph database is "invalid"), the alignment agent is triggered to disambiguate.

[0172] After semantic alignment processing of the orchestration task according to the above rules, the processed orchestration task is sent to the working agent.

[0173] The step of executing the orchestration task by the working intelligent agent based on a pre-trained patent process large model to obtain the task result corresponding to the orchestration task includes:

[0174] The process involves the working agent executing the processed orchestration task based on a pre-trained patent process model to obtain the task result corresponding to the orchestration task.

[0175] By eliminating the ambiguity of terms and nouns among the various agents using the aforementioned patent field ontology, the accuracy of the aligned agents can be improved. Furthermore, by performing the processed orchestration task through the patent process big model, the accuracy of the output results of the patent process big model can be further improved.

[0176] In some embodiments, identifying user questions related to the patent process to obtain user intent includes:

[0177] The user's question is first processed to obtain a question identification result; wherein, the first processing includes natural language understanding processing, key sentence identification, and question entity identification.

[0178] Specifically, users ask questions through the Chat page, the dialogue agent listens to the user's questions, the dialogue agent performs natural language understanding, and performs keyword and phrase recognition. Based on the organization and user entity database trained by the pre-trained large model, entity recognition is performed, such as organization name, user information, specific year information, status information, etc., to obtain the initial question recognition result.

[0179] The initial user intent is obtained by performing intent matching on the problem identification results.

[0180] Initial user intents are obtained by matching intents using a predefined intent library and intent classification model.

[0181] The user intent is determined based on the initial user intent and the first information corresponding to the initial user intent.

[0182] The first information corresponding to the initial user intent includes the first confidence level of the initial user intent and the user's confirmation instruction information for the initial user intent.

[0183] Specifically, slots are filled based on the identified entity information, and historical dialogues are analyzed based on the dialogue context to ensure the continuity and consistency of intent recognition. The next step is to verify and clarify the intent, ensuring that the issue is clarified after passing the confidence assessment (obtaining the first confidence level) and user confirmation, thus ensuring the accuracy of the intent recognition results and generating the intent result (i.e., the user intent).

[0184] In some embodiments of this application, the step of executing the orchestration task by the working intelligent agent according to a pre-trained patent process large model to obtain the task result corresponding to the orchestration task includes:

[0185] The working agent sends task data to the patent process big model according to the orchestration task; the working agent obtains the task result data output by the patent process big model; and the working agent processes the task result data to obtain the task result.

[0186] Specifically, after the working agent observes the scheduling notification from the supervising agent, the working agent executes the current task operation according to the user's intention and feeds back the task result to the supervising agent.

[0187] When an orchestration task in a working agent needs to call the patent process big model, it calls the corresponding API to interact with data according to the application programming interface (API) gateway layer constraint specifications, and applies the results to the execution.

[0188] like Figure 2 As shown, the patent process scheduling and monitoring system also includes a capability agent layer, which includes various working agents for orchestrating task scheduling.

[0189] In an alternative embodiment, such as Figure 2 As shown, the working intelligent agents include a natural language parsing intelligent agent (i.e., NL2SQLAgent), a graph processing intelligent agent (i.e., graphAgent), a risk warning intelligent agent (i.e., risk warningAgent), and a verification intelligent agent (i.e., SQL verificationAgent).

[0190] Among them, the capability agent layer is designed around the existing business to have NL2SQL Agent, SQL verification agent, chart agent and risk warning agent working together. At the same time, in order to achieve integration with the business data layer, a unified API gateway is used to interact with business data and improve the accuracy of model calculation results.

[0191] like Figure 2 As shown, the patent process scheduling and monitoring system also includes a business data layer. The business data layer mainly designs a general API gateway layer to realize data interaction with the Agent. When the NL2SQLAgent generates the final Structured Query Language (SQL), it calls the business data layer through the API to execute it, obtains the result, and returns it to the supervision Agent.

[0192] The business data layer also includes an authorization module, a process query module, a proposal query module, and an organization query module. The authorization module authorizes the orchestration tasks received by the working intelligent agent. After authorization, the orchestration tasks are executed through the patent process big model. The process query module allows for patent process queries, the proposal query module allows for proposal data queries, and the organization query module allows for proposal organization queries.

[0193] The task result data includes chart data and text data; wherein, the chart data includes task result data obtained by processing patent process chart data; the text data includes task result data obtained by processing patent process template data, task result data obtained by processing patent filing organization data, and task result data obtained by processing patent proposal data.

[0194] The step of processing the task result data through the working intelligent agent to obtain the task result includes:

[0195] The text data is processed using the natural language parsing agent (NL2SQLAgent), and the chart data is processed using the chart processing agent (ChartAgent) to obtain the initial task results and risk information during the processing.

[0196] The risk information is monitored using the risk warning agent to obtain monitoring results;

[0197] The initial task result is verified using the verification agent (i.e., SQL verification agent) to obtain the verification result.

[0198] Specifically, the verification agent (i.e., the SQL verification agent) sends SQL verification suggestions to the natural language parsing agent (i.e., the NL2SQL agent), the graph processing agent (i.e., the graph agent), and the risk warning agent (i.e., the risk warning agent), respectively.

[0199] The task result is obtained based on the monitoring result, the verification result, and the initial task result.

[0200] If the monitoring results indicate no abnormalities and the verification results indicate that the verification has passed, the initial task result obtained will be used as the task result; otherwise, the initial task result will be considered incorrect and needs to be obtained again.

[0201] In some embodiments, the method further includes:

[0202] Based on the second confidence level of the working agent and the task result data corresponding to the working agent, the third confidence level of the task result is obtained;

[0203] The accuracy of the task result is determined based on the third confidence level.

[0204] In this embodiment, to improve the quality of user dialogue, after the monitoring agent receives the task result, it evaluates the result to determine its accuracy. If the accuracy indicates a quality problem, it triggers an action to regenerate the result and conduct a quality assessment. To improve the efficiency of risk warning, a confidence-based voting weighted algorithm is proposed. This algorithm combines the results from multiple agents and uses a voting method to calculate the confidence level of the entire agent chain.

[0205] In the above multi-Agent scheduling process, although the terminology is aligned, there may still be inconsistencies or contradictions in the conclusions drawn by multiple agents. In order to improve the accuracy of the results obtained by the multi-Agent scheme, this embodiment proposes a mechanism for fusing multi-Agent voting and confidence during the scheduling process to improve the quality of the dialogue results after multi-Agent orchestration.

[0206] The formula for the confidence-weighted voting model is as follows:

[0207]

[0208] in: Indicates the third confidence level. This represents the initial task result predicted by the i-th working agent. This represents the second confidence level of the i-th working agent, which is determined based on historical accuracy and the model output probability of the agent.

[0209] Specifically, the third confidence level is also updated using a Bayesian update function to dynamically adjust the weights.

[0210] For example, the second confidence level of the risk warning agent is 0.85 and the weight dynamically adjusted by the Bayesian update function is 0.4; the second confidence level of the chart agent is 0.6 and the weight dynamically adjusted by the Bayesian update function is 0.3; and the second confidence level of the SQL verification agent is 0.7 and the weight dynamically adjusted by the Bayesian update function is 0.3.

[0211] When the Coordination Agent receives the task result returned by the Supervision Agent, it transmits the task result to the Dialogue Agent. After observing the result information from the Coordination Agent, the Dialogue Agent transmits the result to the user.

[0212] In summary, the coordination and supervision layer provided in this application embodiment consists of a supervision agent and a coordination agent. The supervision agent is explicitly defined as the core and brain of the system's processes. It is responsible for normal task planning, scheduling, and execution, and possesses the final business decision-making authority. The coordination agent is explicitly defined as the core of exception handling. As a specialized cluster, it provides specific technical fault-tolerance results such as retries, degradation, and circuit breaking, and returns these results as "suggestions" to the supervision agent, which makes the final decision. A reasoning and acting (ReAct) architecture is adopted among the multiple agents to achieve collaborative work, ultimately receiving user intent and obtaining response results.

[0213] In the patent intelligent process query and process intelligent analysis services of this application, data interaction with the business system is required. In order to ensure smooth data interaction between multiple agents and the business system, a unified data interaction specification is provided, which solves the problem of agents using business system data and improves data accuracy and agent development efficiency.

[0214] It should also be noted that the embodiments of this application also provide the specific structure of the working intelligent agent, such as... Figure 3 As shown, the specific structure of the working agent includes four modules: entity management module, data table configuration management module, API scheduling module, and authorization management module, to achieve unified data interaction. The working agent interacts with the unified API gateway and the authorization module. When data interaction is required, it interacts with the authorization module through the authorization management module to obtain a gateway access token. After obtaining the token, it accesses the unified API gateway through the API scheduling module, thereby realizing data interaction with the business system.

[0215] The entity management module serves as the central hub of the business dictionary. It uniformly maintains standard entities in the patent scenario (applicant, inventor, International Patent Classification (IPC), legal status, agency, etc.) and their identifiers (IDs), aliases, encoding mappings, and lifecycle states, ensuring that all agents point to the same business object during queries, statistics, and anomaly detection. It supports hot updates to the entity tree and version rollback, enabling comprehensive updates for any changes made. Furthermore, it integrates with the data table configuration module, providing automatic synonym expansion for NL2SQL and entity-level blacklists and whitelists for the permission module. Simultaneously, its ontology-based semantic alignment mechanism eliminates errors and security issues caused by name inconsistencies or entity drift.

[0216] The data table configuration management module serves as the "foundation". Through the construction of "three databases in one": structure database + business corpus + constraint rule database, vertical knowledge such as enterprise aliases, IPC abbreviations and legal status codes is bound to fields, so as to realize automatic alignment of database values by natural language; meanwhile, in the process of training and generating SQL, the data table configuration management module provides basic semantic alignment, such as converting natural language into database table names and table fields.

[0217] In addition to implementing unified authentication, the API scheduling module is mainly responsible for unified scheduling of all business system services and unified packaging of results. When an error occurs, it is responsible for implementing error handling and providing a corresponding retry mechanism to ensure that users' questions can get correct data responses.

[0218] The authorization management module completes the authorization of the working Agent by the business system. After the working Agent obtains the authorization, it calls the unified gateway service. After the unified gateway service obtains the authorization information, it performs authentication with the business authorization module. If the authentication is passed, the corresponding business API is allowed to be accessed; otherwise, a message indicating no permission is returned.

[0219] Hereinafter, a specific embodiment is used to illustrate the specific process of the scheduling monitoring method for the patent process provided by the present application:

[0220] The user's question is to query the total number of proposals that have passed internal review of Company A;

[0221] Processing process:

[0222] (1) After the conversation Agent obtains the user's question, it performs intent recognition to obtain the user's intent. In this process, semantic analysis is performed in combination with the context, and proper nouns such as internal review, pass and proposal are aligned through the ontology library;

[0223] (2) The user intent is submitted to the coordination Agent, which automatically arranges services according to the user intent, and the call chain is NL2SQLAgent, SQL verification Agent, and risk Agent;

[0224] (3) After being scheduled to the NL2SQLAgent, the final statement generated is: select count(*) from patent where year=2024 and unit_name like ‘%Company A%’ and status=1;

[0225] (4) The supervising Agent delivers the calculation result to the next Agent: the SQL verification Agent;

[0226] (5) After the verification result is returned, the supervising Agent schedules the business data module to execute the SQL and obtain the statistical result;

[0227] (6) After the risk agent assessment, return the coordinated agent;

[0228] (7) After obtaining the data results, the Coordination Agent returns the results to the Dialogue Agent;

[0229] (8) The dialogue agent displays the results to the user: 247.

[0230] This application's embodiments are based on a large model, which trains on patent process templates, organizational data, proposal data, and chart data to generate an intelligent application specifically designed to assist in patent process review. This model addresses the problem of overly complex process paths, allowing users to quickly query and process pending tasks through the large model's chat mode, while simultaneously monitoring process progress in real time.

[0231] Furthermore, this application's embodiments can also schedule and orchestrate patent processes based on a multi-Agent ReAct collaborative mode, enabling functions such as process progress query, process monitoring and analysis query, and intelligent generation of analysis reports; it can monitor the status of process instances and process tasks in real time, use model inference to detect abnormal process instances, and thus quickly issue process processing anomaly alarms, improving process approval efficiency. A general API for Agent-Enterprise data interaction is specifically designed to solve the problem of Agent calling enterprise data, improving the accuracy of data application and Agent development efficiency.

[0232] This application implements an intelligent process design module based on the Agent-based ReAct architecture, enabling patent process progress query, process monitoring and analysis, and abnormal process analysis. In this process, a large-scale model-based NL2SQL model is used to convert natural language into SQL. Simultaneously, multiple Agents are introduced based on different current business needs. Combining reasoning and action capabilities, Agents solve problems in the patent process by continuously thinking, acting, and observing results. This allows each Agent based on the large patent model to more flexibly handle complex tasks and can be enhanced by integrating with patent business systems and data reporting tools.

[0233] This application proposes a Patent Ontology Alignment (POA) technology to achieve semantic alignment and conflict resolution mechanisms among agents, thereby improving the consistency of results obtained by multiple agents. A dual-branch structure is adopted, using ViT to encode the spatial features of patent figures and LLM to encode the text description sequence. A cross-attention mechanism is used to achieve cross-modal alignment between images and text. During the training phase, a dynamic weight allocation strategy is introduced, adjusting the gradient contribution ratio of ViT and LLM in real time based on image-text relevance, solving the modal imbalance problem caused by traditional fixed weights, improving the accuracy of patent figure recognition, and reducing the localization error of key features in the claims. Through a unified API gateway and a unified authentication scheme, authentication and API scheduling submodules are built into the agents, enabling rapid fulfillment of agent access to business data requirements. This also simplifies the requirements for different business access agents, solving the problems of difficult agent-business data integration and high authentication complexity, thereby improving the accuracy and versatility of agents in enterprise applications. By using a confidence-based voting weighted model, confidence scores are calculated for the content returned by multiple agents. By scheduling agents based on this confidence score, the quality of the generated results can be determined, thus improving the accuracy of the content generated by the ReAct architecture and alleviating the drift problem in large models.

[0234] This application can solve multiple problems in the patent management process by constructing a large patent model and entity library. Many model-based business requirements do not need to be redeveloped. Only multiple agents need to be built to solve the business requirements, thereby greatly improving the efficiency of business requirement development and delivery and reducing costs.

[0235] This application solves the problem of complex development in current business by decomposing complex intelligent requirements through the ReAct multi-Agent architecture. It enables intelligent business to be implemented through agent task orchestration, thereby reducing the complexity of intelligent development, improving development efficiency, and making the overall architecture highly scalable and easy to promote.

[0236] This application addresses the current need for visualization in process management by introducing a data reporting agent. It enables users to convert data into data analysis reports through a question-and-answer format, thereby resolving the pain point of data visualization in user management.

[0237] This application provides a general solution for agent-business system data interaction, which solves the complexity problem of agent-business system interaction, and also solves the data risk problem of agent direct connection to database, simplifies the complexity of new business access, and improves agent scalability.

[0238] like Figure 4 As shown in the embodiments of this application, a scheduling and monitoring system for a patent process is also provided, including:

[0239] Dialogue agent 401, which is used to identify user questions regarding the patent process and obtain user intent;

[0240] A coordinating and supervising intelligent agent 402 is used to decompose and orchestrate the user intent to obtain an orchestration task for the patent process and a working intelligent agent corresponding to the orchestration task; wherein, the orchestration task includes a patent process query task and / or a patent process monitoring task.

[0241] A working intelligent agent 403 is used to execute the orchestration task according to a pre-trained patent process big model to obtain the task result corresponding to the orchestration task; wherein, the patent process big model is used to output the query result of the patent process according to the query task of the patent process and / or output the supervision result of the patent process according to the monitoring task of the patent process.

[0242] Optionally, the system further includes:

[0243] The data acquisition module is used to acquire training data for the patent process, wherein the training data includes: patent process template data, data of organizations that file patents, patent proposal data, and patent process chart data.

[0244] The first processing module is used to generate training targets based on the patent process query task and / or the patent process monitoring task.

[0245] The second processing module is used to train the initial model for querying patent processes and / or monitoring patent processes based on the training data and the training objective, so as to obtain the large model of the patent process.

[0246] Optionally, the initial model includes a transformation encoder and a recurrent neural network;

[0247] The second processing module is specifically used for:

[0248] The patent flowchart data is processed using a conversion encoder to extract visual features, resulting in image features.

[0249] The patent process template data is subjected to text feature processing to obtain the first text feature;

[0250] The organizational data of the patent-claiming organization is processed to obtain a second text feature;

[0251] The patent proposal data is processed to obtain a third text feature;

[0252] Using the recurrent neural network, the initial model is trained based on the training objective, the image features, the first text features, the second text features, and the third text features to obtain the large model of the patent process.

[0253] Optionally, the second processing module is specifically used for:

[0254] The patent flowchart data is segmented to obtain chart data blocks;

[0255] The chart data blocks are linearly mapped and flattened to obtain a mapping vector;

[0256] Add positional encoding to the mapping vector to obtain a sequence of chart data blocks;

[0257] The image features are obtained by processing the sequence of chart data blocks using the multi-head self-attention mechanism and feedforward neural network in the transformation encoder.

[0258] Optionally, the second processing module is specifically used for:

[0259] The image features, the first text features, the second text features, and the third text features are fused together to obtain fused features.

[0260] The recurrent neural network is used to obtain the temporal dependency of the fused features and the contextual information of the fused features;

[0261] By utilizing a bidirectional cross-attention mechanism, visual features for text perception and visually guided text features are obtained based on the temporal dependency of the fused features and the contextual information of the fused features.

[0262] A first scaling factor and a second scaling factor are obtained based on the similarity between the image features and the text features using a dynamic gradient gating mechanism; wherein the text features include the first text feature, the second text feature, and the third text feature;

[0263] The target text sequence is obtained based on the first scaling factor, the visual features of the text perception, the second scaling factor, and the visually guided text features;

[0264] Based on the loss information between the training objective and the target text sequence, the model parameters of the initial model are updated to obtain the large model of the patent process.

[0265] Optionally, the coordinating and supervising intelligent agent 402 is specifically used for:

[0266] The orchestration task is semantically decomposed to obtain at least one level of task statements;

[0267] The task statements at at least one level are semantically aligned according to preset semantic alignment rules to obtain the processed orchestration task; wherein, the preset semantic alignment rules include the association between task statements at different levels and the preset standard statements corresponding to each level of task statements.

[0268] Send the processed orchestration task to the working intelligent agent;

[0269] Specifically, the working intelligent agent 403 is used for:

[0270] The process involves the working agent executing the processed orchestration task based on a pre-trained patent process model to obtain the task result corresponding to the orchestration task.

[0271] Optionally, the dialogue agent 401 is specifically used for:

[0272] The user's question undergoes a first processing step to obtain a question identification result; wherein, the first processing step includes natural language understanding processing, key sentence identification, and question entity identification;

[0273] The initial user intent is obtained by performing intent matching on the problem identification results.

[0274] The user intent is determined based on the initial user intent and the first information corresponding to the initial user intent;

[0275] The first information corresponding to the initial user intent includes the first confidence level of the initial user intent and the user's confirmation instruction information for the initial user intent.

[0276] Optionally, the working intelligent agent 403 is specifically used for:

[0277] The working intelligent agent sends task data to the patent process big model according to the orchestrated tasks;

[0278] The task result data output by the large model of the patent process is obtained through the working intelligent agent;

[0279] The task result is obtained by processing the task result data through the working intelligent agent.

[0280] Optionally, the working intelligent agent includes a natural language parsing intelligent agent, a graph processing intelligent agent, a risk warning intelligent agent, and a verification intelligent agent;

[0281] The task result data includes chart data and text data;

[0282] The natural language parsing agent is used to process the text data, and the graph processing agent is used to process the graph data to obtain the initial task results and risk information during the processing.

[0283] The risk warning intelligent agent is used to monitor the risk information and obtain monitoring results;

[0284] The verification agent is used to verify the initial task result and obtain the verification result; the task result is obtained based on the monitoring result, the verification result and the initial task result.

[0285] Optionally, the risk warning intelligent agent is also used for:

[0286] Based on the second confidence level of the working agent and the task result data corresponding to the working agent, the third confidence level of the task result is obtained;

[0287] The accuracy of the task result is determined based on the third confidence level.

[0288] It should be noted that the patent process scheduling and monitoring device provided in this application embodiment is a device capable of executing the above-described patent process scheduling and monitoring method. Therefore, all embodiments of the above-described patent process scheduling and monitoring method are applicable to this device and can achieve the same or similar technical effects.

[0289] like Figure 5 As shown in the figure, this application embodiment also provides a patent process scheduling and monitoring device, including: a processor 501; and a memory 503 connected to the processor 501 via a bus interface 502. The memory 503 is used to store the programs and data used by the processor 501 when performing operations, and the processor 501 calls and executes the programs and data stored in the memory 503.

[0290] The transceiver 504 is connected to the bus interface 502 and is used to receive and send data under the control of the processor 501. Specifically, the processor 501 is used to read the program in the memory 503 and to execute the following processes:

[0291] Identify user questions related to the patent process to obtain user intent;

[0292] The user intent is decomposed and orchestrated to obtain orchestration tasks for the patent process and corresponding working agents for the orchestration tasks; wherein, the orchestration tasks include patent process query tasks and / or patent process monitoring tasks.

[0293] The working agent executes the orchestration task according to the pre-trained patent process big model to obtain the task result corresponding to the orchestration task; wherein, the patent process big model is used to output the query result of the patent process according to the query task of the patent process and / or output the supervision result of the patent process according to the monitoring task of the patent process.

[0294] Optionally, the processor 501 is further configured to:

[0295] Acquire training data for the patent process, wherein the training data includes: patent process template data, data of organizations that file patents, patent proposal data, and patent process chart data;

[0296] Training targets are generated based on the patent process query task and / or patent process monitoring task.

[0297] Based on the training data and the training objective, the initial model used for querying patent processes and / or for monitoring patent processes is trained to obtain the large patent process model.

[0298] Optionally, the initial model includes a transformation encoder and a recurrent neural network;

[0299] Specifically, the processor 501 is used for:

[0300] The patent flowchart data is processed using a conversion encoder to extract visual features, resulting in image features.

[0301] The patent process template data is subjected to text feature processing to obtain the first text feature;

[0302] The organizational data of the patent-claiming organization is processed to obtain a second text feature;

[0303] The patent proposal data is processed to obtain a third text feature;

[0304] Using the recurrent neural network, the initial model is trained based on the training objective, the image features, the first text features, the second text features, and the third text features to obtain the large model of the patent process.

[0305] Optionally, the processor 501 is specifically used for:

[0306] The patent flowchart data is segmented to obtain chart data blocks;

[0307] The chart data blocks are linearly mapped and flattened to obtain a mapping vector;

[0308] Add positional encoding to the mapping vector to obtain a sequence of chart data blocks;

[0309] The image features are obtained by processing the sequence of chart data blocks using the multi-head self-attention mechanism and feedforward neural network in the transformation encoder.

[0310] Optionally, the processor 501 is specifically used for:

[0311] The image features, the first text features, the second text features, and the third text features are fused together to obtain fused features.

[0312] The recurrent neural network is used to obtain the temporal dependency of the fused features and the contextual information of the fused features;

[0313] By utilizing a bidirectional cross-attention mechanism, visual features for text perception and visually guided text features are obtained based on the temporal dependency of the fused features and the contextual information of the fused features.

[0314] A first scaling factor and a second scaling factor are obtained based on the similarity between the image features and the text features using a dynamic gradient gating mechanism; wherein the text features include the first text feature, the second text feature, and the third text feature;

[0315] The target text sequence is obtained based on the first scaling factor, the visual features of the text perception, the second scaling factor, and the visually guided text features;

[0316] Based on the loss information between the training objective and the target text sequence, the model parameters of the initial model are updated to obtain the large model of the patent process.

[0317] Optionally, the processor 501 is further configured to:

[0318] The orchestration task is semantically decomposed to obtain at least one level of task statements;

[0319] The task statements at at least one level are semantically aligned according to preset semantic alignment rules to obtain the processed orchestration task; wherein, the preset semantic alignment rules include the association between task statements at different levels and the preset standard statements corresponding to each level of task statements.

[0320] Send the processed orchestration task to the working intelligent agent;

[0321] Specifically, the processor 501 is used for:

[0322] The process involves the working agent executing the processed orchestration task based on a pre-trained patent process model to obtain the task result corresponding to the orchestration task.

[0323] Optionally, the processor 501 is specifically used for:

[0324] The user's question undergoes a first processing step to obtain a question identification result; wherein, the first processing step includes natural language understanding processing, key sentence identification, and question entity identification;

[0325] The initial user intent is obtained by performing intent matching on the problem identification results.

[0326] The user intent is determined based on the initial user intent and the first information corresponding to the initial user intent;

[0327] The first information corresponding to the initial user intent includes the first confidence level of the initial user intent and the user's confirmation instruction information for the initial user intent.

[0328] Optionally, the processor 501 is specifically used for:

[0329] The working intelligent agent sends task data to the patent process big model according to the orchestrated tasks;

[0330] The task result data output by the large model of the patent process is obtained through the working intelligent agent;

[0331] The task result is obtained by processing the task result data through the working intelligent agent.

[0332] Optionally, the working intelligent agent includes a natural language parsing intelligent agent, a graph processing intelligent agent, a risk warning intelligent agent, and a verification intelligent agent;

[0333] The task result data includes chart data and text data;

[0334] Specifically, the processor 501 is used for:

[0335] The text data is processed using the natural language parsing agent, and the chart data is processed using the chart processing agent to obtain the initial task results and risk information during the processing.

[0336] The risk warning intelligent agent is used to monitor the risk information and obtain monitoring results;

[0337] The verification agent is used to verify the initial task result to obtain the verification result;

[0338] The task result is obtained based on the monitoring result, the verification result, and the initial task result.

[0339] Optionally, the processor 501 is further configured to:

[0340] Based on the second confidence level of the working agent and the task result data corresponding to the working agent, the third confidence level of the task result is obtained;

[0341] The accuracy of the task result is determined based on the third confidence level.

[0342] Among them, Figure 5 In this context, the bus architecture may include any number of interconnected buses and bridges, specifically linking various circuits together, represented by one or more processors (processor 501) and memory (memory 503). The bus architecture may also link various other circuits such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and therefore will not be described further herein. A bus interface provides a user interface 505. A transceiver 504 may be multiple elements, including transmitters and receivers, providing units for communicating with various other devices over a transmission medium. Processor 501 is responsible for managing the bus architecture and general processing, and memory 503 may store data used by processor 501 during operation.

[0343] In addition, specific embodiments of this application also provide a readable storage medium storing a computer program thereon, wherein when the program is executed by a processor, it implements the steps in the scheduling and monitoring method of the patent process as described above.

[0344] In the several embodiments provided in this application, it should be understood that the disclosed methods and apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0345] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can be physically included separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0346] The integrated units implemented as software functional units described above can be stored in a computer-readable storage medium. These software functional units, stored in a storage medium, include several instructions that cause a computer device (which may be a personal computer, server, or network device, etc.) to execute partial steps of the resource selection method described in the various embodiments of this application, or to execute partial steps of the information transmission method described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0347] A specific embodiment of this application also provides a computer program product, including computer instructions, which, when executed by a processor, implement the above-described functionality. Figure 1 The various processes of the method embodiments shown can achieve the same technical effect, and will not be described again here to avoid repetition.

[0348] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A method for scheduling and monitoring a patented process, characterized in that, include: Identify user questions related to the patent process to obtain user intent; The user intent is decomposed and orchestrated to obtain orchestration tasks for the patent process and corresponding working agents for the orchestration tasks; wherein, the orchestration tasks include patent process query tasks and / or patent process monitoring tasks. The working agent executes the orchestration task according to the pre-trained patent process big model to obtain the task result corresponding to the orchestration task; wherein, the patent process big model is used to output the query result of the patent process according to the query task of the patent process and / or output the supervision result of the patent process according to the monitoring task of the patent process.

2. The method according to claim 1, characterized in that, The method further includes: Acquire training data for the patent process, wherein the training data includes: patent process template data, data of organizations that file patents, patent proposal data, and patent process chart data; Training targets are generated based on the patent process query task and / or patent process monitoring task. Based on the training data and the training objective, the initial model used for querying patent processes and / or for monitoring patent processes is trained to obtain the large patent process model.

3. The method according to claim 2, characterized in that, The initial model includes a transformation encoder and a recurrent neural network; Specifically, based on the training data and the training objective, an initial model for querying and / or monitoring patent processes is trained to obtain the large-scale patent process model, including: The patent flowchart data is processed using a conversion encoder to extract visual features, thereby obtaining image features. The patent process template data is subjected to text feature processing to obtain the first text feature; Text feature processing is performed on the organizational data of the patent-claiming organization to obtain a second text feature; The patent proposal data is processed to obtain a third text feature; Using the recurrent neural network, the initial model is trained based on the training objective, the image features, the first text features, the second text features, and the third text features to obtain the large model of the patent process.

4. The method according to claim 3, characterized in that, The step of using a transformation encoder to perform visual feature extraction processing on the patent flowchart data to obtain image features includes: The patent flowchart data is segmented to obtain chart data blocks; The chart data blocks are linearly mapped and flattened to obtain a mapping vector; Add positional encoding to the mapping vector to obtain a sequence of chart data blocks; The image features are obtained by processing the sequence of chart data blocks using the multi-head self-attention mechanism and feedforward neural network in the transformation encoder.

5. The method according to claim 3, characterized in that, The process of using the recurrent neural network to train the initial model based on the training objective, the image features, the first text features, the second text features, and the third text features to obtain the large-scale patent process model includes: The image features, the first text features, the second text features, and the third text features are fused together to obtain fused features. The recurrent neural network is used to obtain the temporal dependency of the fused features and the contextual information of the fused features; By utilizing a bidirectional cross-attention mechanism, visual features for text perception and visually guided text features are obtained based on the temporal dependency of the fused features and the contextual information of the fused features. A first scaling factor and a second scaling factor are obtained based on the similarity between the image features and the text features using a dynamic gradient gating mechanism; wherein the text features include the first text feature, the second text feature, and the third text feature; The target text sequence is obtained based on the first scaling factor, the visual features of the text perception, the second scaling factor, and the visually guided text features; Based on the loss information between the training objective and the target text sequence, the model parameters of the initial model are updated to obtain the large model of the patent process.

6. The method according to claim 1, characterized in that, After decomposing and orchestrating the user intent to obtain the orchestration task for the patent process and the corresponding working agent for the orchestration task, the method further includes: The orchestration task is semantically decomposed to obtain at least one level of task statements; The task statements at at least one level are semantically aligned according to preset semantic alignment rules to obtain the processed orchestration task; wherein, the preset semantic alignment rules include the association between task statements at different levels and the preset standard statements corresponding to each level of task statements. Send the processed orchestration task to the working intelligent agent; The step of executing the orchestration task by the working intelligent agent based on a pre-trained patent process large model to obtain the task result corresponding to the orchestration task includes: The process involves the working agent executing the processed orchestration task based on a pre-trained patent process model to obtain the task result corresponding to the orchestration task.

7. The method according to claim 1, characterized in that, The process of identifying user questions related to the patent process and obtaining user intent includes: The user's question undergoes a first processing step to obtain a question identification result; wherein, the first processing step includes natural language understanding processing, key sentence identification, and question entity identification; The initial user intent is obtained by performing intent matching on the problem identification results. The user intent is determined based on the initial user intent and the first information corresponding to the initial user intent; The first information corresponding to the initial user intent includes the first confidence level of the initial user intent and the user's confirmation instruction information for the initial user intent.

8. The method according to claim 1, characterized in that, The step of executing the orchestration task through the working intelligent agent based on a pre-trained patent process large model to obtain the task result corresponding to the orchestration task includes: The working intelligent agent sends task data to the patent process big model according to the orchestrated tasks; The working intelligent agent obtains the task result data output by the large model of the patent process. The task result is obtained by processing the task result data through the working intelligent agent.

9. The method according to claim 8, characterized in that, The working intelligent agents include a natural language parsing intelligent agent, a graph processing intelligent agent, a risk warning intelligent agent, and a verification intelligent agent; The task result data includes chart data and text data; The step of processing the task result data through the working intelligent agent to obtain the task result includes: The text data is processed using the natural language parsing agent, and the chart data is processed using the chart processing agent to obtain the initial task results and risk information during the processing. The risk warning intelligent agent is used to monitor the risk information and obtain monitoring results; The verification agent is used to verify the initial task result to obtain the verification result; The task result is obtained based on the monitoring result, the verification result, and the initial task result.

10. The method according to claim 1, characterized in that, The method further includes: Based on the second confidence level of the working agent and the task result data corresponding to the working agent, the third confidence level of the task result is obtained; The accuracy of the task result is determined based on the third confidence level.

11. A scheduling and monitoring system for a patented process, characterized in that, include: A dialogue agent, which is used to identify user questions regarding the patent process and obtain user intent; A coordinating and supervising intelligent agent is used to decompose and orchestrate the user intent to obtain an orchestration task for the patent process and a working intelligent agent corresponding to the orchestration task; wherein, the orchestration task includes a patent process query task and / or a patent process monitoring task. A working intelligent agent is used to execute the orchestration task according to a pre-trained patent process big model to obtain the task result corresponding to the orchestration task; wherein, the patent process big model is used to output the query result of the patent process according to the query task of the patent process and / or output the supervision result of the patent process according to the monitoring task of the patent process.

12. A patented process scheduling and monitoring device, characterized in that, include: A processor, a memory, and a program stored in the memory and executable on the processor, wherein the program, when executed by the processor, implements the steps of the scheduling and monitoring method of the patent process as described in any one of claims 1 to 10.

13. A readable storage medium, characterized in that, The readable storage medium stores a program that, when executed by a processor, implements the steps of the scheduling and monitoring method of the patent process as described in any one of claims 1 to 10.

14. A computer program product, characterized in that, It includes computer instructions that, when executed by a processor, implement the steps in the scheduling and monitoring method of the patent process as described in any one of claims 1 to 10.

Images