Patent disclosure book auxiliary writing method and system based on multi-agent cooperation

Abstract

The invention discloses a patent disclosure auxiliary writing method and system based on multi-agent collaboration. The method comprises the steps that a multi-agent system composed of a collaboration agent, an agent agent and an expert agent is constructed, a knowledge base and a decision module of each agent are configured, and a collaboration architecture of a task scheduler, an arbiter and a shared information space is established; the method comprises the following steps: receiving an original technical scheme of a user, and constructing a structured intermediate representation through multi-agent parallel semantic analysis and feature extraction; based on the representation, a writing task sequence is generated through dynamic task decomposition, intelligent agent division is coordinated through a contract network protocol, and content generation and verification are completed based on a multi-round debate mechanism; performing multi-dimensional consistency detection on the generated draft, and eliminating logic conflicts and expressing defects through iterative revision; and finally, an interest portrait is generated based on the user technical scheme and the optimized draft, interest retrieval is performed, a technical content report is output, and the efficiency and quality of disclosure book writing are improved.

Description

Technical Field

[0001] This invention relates to the field of text generation assistance technology, and in particular to a method and system for assisting in the writing of patent disclosure documents based on multi-agent collaboration. Background Technology

[0002] Patent disclosures serve as a crucial bridge connecting technological innovation and legal protection; their quality directly impacts the success of a patent application and the scope of protection. Currently, the drafting of patent disclosures primarily relies on manual collaboration between inventors and patent attorneys. Inventors provide the original technical solutions, but their descriptions often focus on the technical implementation itself, resulting in issues such as non-standard terminology, unclear logic, and a lack of patent law perspective. Patent attorneys, on the other hand, must expend considerable time and effort to understand, organize, and translate this technical information, supplementing necessary legal elements and reconstructing it according to strict patent law norms. This process is highly dependent on the attorney's personal experience and professional expertise, leading to inefficiency and risks such as omissions of key technical points and improper definition of the scope of protection due to poor communication or misunderstandings.

[0003] In recent years, although some research has attempted to utilize artificial intelligence technology to assist in drafting, existing solutions are mostly based on single-agent models, resulting in functional limitations. For example, some tools can only perform text format checks or simple terminology standardization, failing to conduct substantive analysis of the depth and innovativeness of technical solutions; others focus on matching legal clauses, lacking the ability to understand complex technical logic. This "lone wolf" approach makes it difficult to simultaneously ensure the accuracy of technical descriptions and the standardization of legal texts, failing to achieve end-to-end intelligent assistance from technical solutions to high-quality disclosure documents. Therefore, there is an urgent need in this field for a solution that can deeply integrate technical understanding and legal knowledge to achieve intelligent collaborative drafting, fundamentally improving the efficiency and quality of disclosure document drafting. Summary of the Invention

[0004] This invention overcomes the shortcomings of the prior art and provides a method and system for assisting in the writing of patent disclosure documents based on multi-agent collaboration. Its main purpose is to improve the efficiency and quality of patent disclosure document writing.

[0005] To achieve the above objectives, the first aspect of this invention provides a method for assisting in drafting a patent disclosure based on multi-agent collaboration, comprising: Establish a multi-agent writing assistance system consisting of collaborative agents, agent agents, and expert agents, and configure dedicated knowledge bases and decision processing modules for each agent. Construct a collaborative processing architecture with a task scheduler, arbitrator, and shared information space. The system receives the original technical solution description input by the user and inputs it into the multi-agent writing assistance system to call on collaborative multi-agents to perform parallel semantic parsing and feature extraction, and constructs a structured intermediate representation that includes technical problems, technical solutions and technical effects. Based on the structured intermediate representation, a dynamic task decomposition algorithm is used to generate a task sequence for writing the disclosure document. The task allocation and coordination among agents are realized through the contract network protocol. The content generation and cross-validation are completed based on a multi-round debate mechanism to obtain the draft disclosure document. The generated draft disclosure document undergoes multi-dimensional consistency checks, and predefined conflict resolution strategies are used to coordinate ambiguities. Logical conflicts and expression defects are eliminated through an iterative revision process, and an optimized draft disclosure document is generated and pushed out. Based on the original technical solution description, a user interest profile is generated by combining the optimized disclosure draft, and an interest retrieval is performed based on the user interest profile to generate an interest-based technical content report.

[0006] In this solution, the establishment of a multi-agent writing assistance system consisting of collaborative agents, agent agents, and expert agents, and the configuration of dedicated knowledge bases and decision processing modules for each agent, along with the construction of a collaborative processing architecture comprising a task scheduler, an arbitrator, and a shared information space, specifically includes: Deploy independent microservice instances in a distributed computing environment, integrate a workflow engine based on Apache Airflow as a task scheduler, the workflow engine is pre-configured with several standard chapter templates of patent disclosure documents, and the execution flow of the directed acyclic graph is constructed between the chapters through data dependencies, and a collaborative processing architecture of arbitrator, shared information space and task scheduler is constructed to obtain a collaborative intelligent agent; The arbitrator is implemented using a rule engine based on Drools. The loaded conflict resolution rule base includes a set of rules for agent priority, a set of rules for content consistency, and a set of rules for semantic conflict detection. The shared information space is implemented using a MongoDB database. The data pattern of the shared information space includes fields for task ID, generating agent, content version, content data, and status identifier. Patent-related knowledge data is obtained by using big data networks, patent drafting rule data is extracted, and natural language processing technology is used for word segmentation, entity recognition and relation extraction. A patent rule knowledge graph stored in the form of RDF triples is constructed. The decision processing module is configured through the Drools rule engine using the patent rule knowledge graph, and finally the agent intelligent agent is obtained. Technical documents are extracted from the patent-related knowledge data, term weights are calculated using the TF-IDF algorithm, key technical terms are filtered by setting thresholds, a term co-occurrence network is constructed using the TextRank algorithm based on a pre-set co-occurrence window, and clusters of technical concepts are identified using the Louvain community discovery algorithm. A technical knowledge base containing technical term nodes, technical effect nodes, and technical problem nodes is established. Deploy a BERT pre-trained language model and continuously pre-train and fine-tune it using technical documents in the knowledge base of the aforementioned technical field to construct an expert agent; by integrating the collaborative agent, agent agent, and expert agent, and setting a service registration and discovery mechanism and a message queue communication protocol, construct a multi-agent writing assistance system.

[0007] In this solution, the establishment of a multi-agent writing assistance system consisting of collaborative agents, agent agents, and expert agents, and the configuration of dedicated knowledge bases and decision processing modules for each agent, along with the construction of a collaborative processing architecture comprising a task scheduler, an arbitrator, and a shared information space, further includes: The workflow of the task scheduler includes: receiving a user request, parsing the technical solution text, generating an initial directed acyclic graph of tasks based on a preset template; using the Kahn topological sorting algorithm to determine the task execution sequence, asynchronously distributing task instructions through a message queue, and ensuring the reliability of task execution through heartbeat detection and timeout retry mechanisms. The arbitrator's conflict resolution mechanism includes: when a content conflict is detected, a multi-round debate process is initiated. The first round adopts a rapid ruling based on rule priority. If no consensus is reached, a voting mechanism based on confidence weighting is initiated. The debate log records the complete argument tree structure, including claim nodes, evidence nodes, and attack edges. The final ruling result is digitally signed to ensure immutability and is synchronized to each agent in real time. The data management process of the shared information space includes: adopting a version control mechanism based on content hashing and generating content fingerprints using the SHA-256 algorithm; setting up a full-text search interface with combined query modes based on time range, agent type, and task status to ensure complete data traceability and efficient access.

[0008] In this scheme, the original technical solution description received from the user is input into the multi-agent writing assistance system, which calls on collaborative agents to perform parallel semantic parsing and feature extraction to construct a structured intermediate representation containing the technical problem, technical solution, and technical effect. Specifically, this includes: The system obtains the original technical solution description input by the user and inputs it into the multi-agent writing assistance system. The collaborative agent uses a hybrid word segmentation algorithm that combines a patent domain dictionary to perform lexical-level segmentation on the original text, and identifies and replaces sensitive information and proper nouns in the text. The system constructs a lexical dependency tree through a dependency parser and marks part-of-speech tags and syntactic roles to generate a standardized text sequence. The standardized text sequence is distributed in parallel to the agent agent and the expert agent through the RabbitMQ message middleware; the agent agent uses the patent rule knowledge graph to generate a multi-dimensional rule distributed representation through the GraphSAGE graph neural network encoder, adopts a bidirectional LSTM-CRF architecture with 4 hidden layers and combines a multi-head self-attention mechanism to identify entities and limiting phrases in the text, and outputs patent rule feature vectors. Meanwhile, the expert agent captures the contextual information of the technical text through a multi-layer Transformer encoder, extracts technical entities using a pointer network-based BIO annotation sequence entity recognition mechanism, constructs the predicate-argument structure of the operation process through LTP dependency parsing and semantic role annotation, and finally obtains a multi-dimensional technical feature vector. The collaborative intelligent agent performs multi-stage feature alignment and fusion based on the patent rule feature vector and multi-dimensional technical feature vector, performs scale unification processing on the two types of feature vectors based on maximum and minimum normalization, and constructs a feature association matrix by calculating the cosine similarity between feature vectors. The feature association matrix is ​​regarded as the adjacency matrix of a weighted undirected graph. The corresponding degree matrix and Laplacian matrix are calculated. Then, the Laplacian matrix is ​​decomposed into features. The eigenvectors corresponding to the first k smallest eigenvalues ​​are selected to form a new feature space. In the new feature space, the K-means clustering algorithm is used to divide the original features into k different semantic clusters. For each feature cluster, the mean cosine similarity of the feature vectors within the cluster is calculated, and highly overlapping feature groups are identified by using a similarity threshold. Among the identified overlapping feature groups, the feature vector closest to the cluster center is retained as a representative feature, while other redundant feature vectors are removed to generate a filtered feature cluster. A multi-head attention mechanism is introduced to represent the query vector by patent rule feature vectors and key-value multi-dimensional technology feature vectors. Attention weights are calculated and feature fusion is performed to finally output multimodal fused features. Based on the multimodal fusion features, a syntactic dependency tree is constructed. A causal relationship graph with technical problems, technical means, and technical effects as causal paths is constructed through the syntactic dependency tree. The betweenness centrality and eigenvector centrality of each node in the causal relationship graph are used as node importance scores to construct a state space. A near-end policy optimization algorithm is used to perform semantic parsing optimization to output an optimized semantic parse tree. The optimized semantic parse tree is then mapped to a structured intermediate representation containing technical problems, technical solutions, and technical effects.

[0009] In this scheme, based on the structured intermediate representation, a dynamic task decomposition algorithm is used to generate a task sequence for writing the disclosure document. The contract network protocol is used to achieve task allocation and coordination among agents. Furthermore, a multi-round debate mechanism is used to complete content generation and cross-validation, resulting in a draft disclosure document. Specifically, this includes: The collaborative intelligent agent performs dynamic task decomposition based on the structured intermediate representation, extracts the topological ordering relationship between each technical node, uses the critical path analysis method to determine the key sequence of task execution, and maps the entities in the structured intermediate representation and the relationships between entities to writing task nodes based on the chapter structure writing rules of the disclosure document, generating a directed acyclic graph of tasks with execution order. Based on the task directed acyclic graph, a contract network is used to conduct distributed task negotiation. The collaborative agent acts as the manager and encapsulates each subtask in the task directed acyclic graph into a task announcement containing task specifications, quality indicators and time constraints. The announcement is then broadcast to all registered agents through a message middleware. After all the agents receive the bidding information, they calculate the semantic matching degree between their own capability attributes and task requirements, combine real-time load status and historical performance indicators, and use multi-attribute utility theory to calculate the comprehensive bidding value. The collaborative agents conduct multi-objective optimization evaluation of all bids after the bidding deadline, and determine the optimal task-agent allocation mapping table based on the multi-objective optimization algorithm. Tasks are dispatched through an optimal task-agent allocation mapping table. The task-bearing agent generates an initial content draft based on its professional knowledge base. The agent agent is responsible for generating the rule content, and the expert agent is responsible for generating the technical content. The initial content draft is generated and cross-validated, with the technical content being validated by the expert agent and the rule content being validated by the agent agent. When content disputes are detected, the verification agent constructs a questioning framework that includes counterexamples and logical rebuttals, and initiates a debate challenge to the corresponding task agent through a message queue. After obtaining a consensus content draft through repeated debates, the content of the next task node is generated. The coordinating agent manages task execution based on the dependency relationships in the task dependency network graph. After all subtasks have completed debate verification, the collaborative agent synthesizes the final draft content of all subtasks based on the task execution order of the directed acyclic graph of tasks, and finally generates the draft disclosure document.

[0010] In this solution, the process of performing multi-dimensional consistency checks on the generated draft disclosure document, coordinating ambiguities using predefined conflict resolution strategies, eliminating logical conflicts and expression defects through an iterative revision process, and generating an optimized draft disclosure document for submission includes: After generating the draft disclosure document, the collaborative intelligent agent constructs a consistency detection framework based on a multi-level graph structure and initializes the framework. The consistency detection framework includes four detection dimensions: technical logic dimension, writing compliance dimension, terminology consistency dimension, and graphic description matching dimension. The draft disclosure document is input into the initialized consistency detection framework for multi-dimensional parallel detection. The technical logic dimension detection is performed by an expert intelligent agent using a technical reasoning mechanism based on SWRL rules, and a technical logic detection report containing inconsistent node pairs and their confidence levels is generated. The agent performs compliance checks on the drafting process, parsing the draft disclosure document into a drafting element graph structure and using a violation pattern matching algorithm based on subgraph isomorphism to generate a compliance check report. The collaborative agent simultaneously performs terminology consistency and image-text matching checks, constructing a full-text terminology distribution vector space, calculating a terminology context similarity matrix, and combining it with an illustration keypoint matching algorithm based on SIFT feature extraction to generate a comprehensive check report that includes terminology consistency and image-text matching. After completing multi-dimensional parallel detection, the generated detection reports are integrated and a conflict evidence framework is constructed based on DS evidence theory. The conflict evidence framework is used to assign probability values ​​to each detected problem instance and calculate the trust function and likelihood function. Finally, a structured conflict detection report containing problem type, location coordinates, severity level, technical element correlation degree and resolution priority is generated.

[0011] In this solution, the process of performing multi-dimensional consistency checks on the generated draft disclosure document, coordinating ambiguities using predefined conflict resolution strategies, eliminating logical conflicts and expression defects through an iterative revision process, and generating an optimized draft disclosure document for submission also includes: The collaborative intelligent agent uses the TOPSIS multi-attribute decision algorithm to rank conflict problems based on the structured conflict detection report. It sets an objective function that includes the severity of the conflict, the scope of its impact, and the complexity of its processing, determines the priority sequence of problem processing, and obtains a ranked conflict problem sequence. The collaborative intelligent agent then obtains conflict resolution strategies in the shared information space through the ranked conflict problem sequence. For technical feature description conflicts, semantic similarity between different expressions is calculated, and word vector models are used to generate modification schemes for description conflicts; for drafting compliance conflicts, the agent's intelligent agent is invoked to generate drafting modification schemes based on the patent rule knowledge graph; for logical chain conflicts, the expert intelligent agent is invoked to monitor and generate logical connection modification schemes in the technical field knowledge base; finally, a set of conflict resolution strategies is generated. The conflict resolution priority is extracted by sorting the conflict problem sequence, and the draft disclosure document is iteratively revised in combination with the conflict resolution strategy set. In each iteration, a consistency check is performed again based on the generated new draft disclosure document. If there are conflict problems that do not achieve the expected resolution effect, the corresponding resolution strategy is marked as an inefficient strategy and a new resolution strategy is obtained. Through an iterative process of repeated detection, revision, and verification until the resolution of all conflicts reaches a preset quality threshold, the final optimized draft disclosure document is output and pushed out.

[0012] In this solution, the step of generating a user interest profile based on the original technical solution description and the optimized draft disclosure document, performing interest retrieval based on the user interest profile, and generating an interest-related technical content report specifically includes: The original technical solution description of the user is extracted from the shared information space and the optimized disclosure draft is obtained. The original technical solution description is processed by word segmentation and dependency parsing based on the patent field dictionary to identify the core technical problem statement and the description of innovation points. At the same time, the optimized disclosure draft is subjected to structured parsing to extract key technical terms, technical problem focus and technical effect features to form an initial set of user interest features. The initial set of user interest features is input into the expert agent. The technical knowledge base is used to perform semantic expansion and association analysis on each user interest feature. Potential related technical concepts are identified by the semantic similarity between technical terms. The technical field distribution corresponding to the original technical solution is extracted by LDA topic analysis. Finally, a multi-dimensional interest feature vector containing core interest features, related technical concepts and technical field preferences is constructed. Based on the multidimensional interest feature vector, user interest features are integrated according to four dimensions: technical field, technical problem, technical means and technical effect, and a user interest profile is constructed. A weight value is set for each dimension, wherein the weight value is determined by the frequency of the corresponding interest feature in the original technical solution and the optimization disclosure document. The user interest profile is converted into a retrieval query vector, and a big data network is introduced to retrieve interest documents through semantic similarity calculation. The retrieval results are sorted by calculating the semantic relevance between the query vector and the document content, and interest documents within a preset range are selected to generate an interest retrieval result set. The interest retrieval result set is imported into an expert intelligent agent for analysis. The technical description features corresponding to each interest document are extracted using key phrase extraction technology to construct an interest summary for each interest document. This summary is then compared with the original technical solution description to obtain the difference analysis results. Finally, an interest technology content report is generated and pushed to the target users.

[0013] The second aspect of the present invention provides a patent disclosure document drafting assistance system based on multi-agent collaboration, the system comprising: a user interaction module, a collaborative agent module, an agent agent module, an expert agent module, a data storage module, and an optimization control module; The user interaction module is used to handle user input and output interaction, including receiving the initial technical solution description, displaying the generated results, and collecting user feedback. It is implemented through a graphical user interface or API interface, supports natural language input, and updates the progress status in real time. The collaborative intelligent agent module, as the central coordinator of the system, is responsible for parsing user input, assigning tasks and driving iterative optimization. It uses natural language processing technology to extract key information, calls other intelligent agent modules, and compares the output content to generate a difference report, so as to achieve multi-round interaction and feedback integration. The agent intelligent agent module is used for the standardized writing of patent disclosure documents. It has a built-in patent rule knowledge graph and template engine, and generates a text structure that conforms to the patent format through rule matching and grammatical analysis. The expert agent module is responsible for improving and supplementing the technical solution, integrating machine learning models and technical knowledge bases, identifying missing points in the technical solution through semantic analysis, and recommending technical directions of interest. The data storage module is used to manage the system's persistent data, including rule base, knowledge base, user history records and temporary cache. It stores and retrieves data through a database system, supports efficient querying and updating, and ensures the consistency of data sharing among modules. The optimization control module is used to invoke innovative optimization suggestions, requiring the agent to continue optimizing the technical solution.

[0014] A third aspect of the present invention provides a computer-readable storage medium comprising a method program for assisting in drafting a patent disclosure document based on multi-agent collaboration. When the method program is executed by a processor, it implements the steps of the method program for assisting in drafting a patent disclosure document based on multi-agent collaboration as described in any of the preceding claims. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments or examples of the present invention, the drawings used in the embodiments or examples will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained according to these drawings without creative effort.

[0016] Figure 1 A flowchart of the first method for assisting in drafting a patent disclosure document based on multi-agent collaboration, provided in an embodiment of the present invention; Figure 2 A flowchart of a second method for assisting in drafting a patent disclosure document based on multi-agent collaboration, provided as an embodiment of the present invention; Figure 3 This is a block diagram of a patent disclosure document drafting assistance system based on multi-agent collaboration, provided in one embodiment of the present invention. The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0017] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0018] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0019] Figure 1 A flowchart of the first method for assisting in drafting a patent disclosure document based on multi-agent collaboration, provided in an embodiment of the present invention; like Figure 1 As shown, the present invention provides a first method flowchart for a patent disclosure document drafting assistance method based on multi-agent collaboration, comprising: S102, establish a multi-agent writing assistance system consisting of collaborative agents, agent agents, and expert agents, and configure dedicated knowledge bases and decision processing modules for each agent, and construct a collaborative processing architecture of task scheduler, arbitrator, and shared information space; S104, Receive the original technical solution description input by the user, input it into the multi-agent writing assistance system to call the collaborative multi-agent to perform parallel semantic parsing and feature extraction, and construct a structured intermediate representation containing technical problems, technical solutions and technical effects; S106, Based on the structured intermediate representation, a dynamic task decomposition algorithm is used to generate a task sequence for writing the disclosure document. The task allocation and coordination among agents are realized through the contract network protocol. The content generation and cross-validation are completed based on a multi-round debate mechanism to obtain a draft disclosure document. S108, Perform multi-dimensional consistency checks on the generated draft disclosure document, use predefined conflict resolution strategies to coordinate ambiguities, eliminate logical conflicts and expression defects through iterative revision process, and generate an optimized draft disclosure document for submission; S110, Based on the original technical solution description and combined with the optimized disclosure draft, generate a user interest profile, perform interest retrieval based on the user interest profile, and generate an interest technology content report.

[0020] Furthermore, in a preferred embodiment of the present invention, the establishment of a multi-agent writing assistance system composed of a collaborative agent, an agent agent, and an expert agent, and the configuration of a dedicated knowledge base and decision processing module for each agent, and the construction of a collaborative processing architecture consisting of a task scheduler, an arbitrator, and a shared information space, specifically includes: Deploy independent microservice instances in a distributed computing environment, integrate a workflow engine based on Apache Airflow as a task scheduler, the workflow engine is pre-configured with several standard chapter templates of patent disclosure documents, and the execution flow of the directed acyclic graph is constructed between the chapters through data dependencies, and a collaborative processing architecture of arbitrator, shared information space and task scheduler is constructed to obtain a collaborative intelligent agent; The arbitrator is implemented using a rule engine based on Drools. The loaded conflict resolution rule base includes a set of rules for agent priority, a set of rules for content consistency, and a set of rules for semantic conflict detection. The shared information space is implemented using a MongoDB database. The data pattern of the shared information space includes fields for task ID, generating agent, content version, content data, and status identifier. Patent-related knowledge data is obtained by using big data networks, patent drafting rule data is extracted, and natural language processing technology is used for word segmentation, entity recognition and relation extraction. A patent rule knowledge graph stored in the form of RDF triples is constructed. The decision processing module is configured through the Drools rule engine using the patent rule knowledge graph, and finally the agent intelligent agent is obtained. Technical documents are extracted from the patent-related knowledge data, term weights are calculated using the TF-IDF algorithm, key technical terms are filtered by setting thresholds, a term co-occurrence network is constructed using the TextRank algorithm based on a pre-set co-occurrence window, and clusters of technical concepts are identified using the Louvain community discovery algorithm. A technical knowledge base containing technical term nodes, technical effect nodes, and technical problem nodes is established. Deploy a BERT pre-trained language model and continuously pre-train and fine-tune it using technical documents in the knowledge base of the aforementioned technical field to construct an expert agent; by integrating the collaborative agent, agent agent, and expert agent, and setting a service registration and discovery mechanism and a message queue communication protocol, construct a multi-agent writing assistance system.

[0021] It should be noted that independent microservice instances are deployed in the distributed computing environment, and a task scheduler is built by integrating a workflow engine based on Apache Airflow. This engine is pre-configured with standard chapter templates from the patent disclosure, and the execution flow of each chapter is constructed based on data dependencies, forming a complete task scheduling system. Simultaneously, an arbitrator and a shared information space are built to form a collaborative processing architecture. The arbitrator is implemented using a rule engine based on Drools, and the loaded conflict resolution rule base includes a set of rules for agent priority, content consistency, and semantic conflict detection, providing decision support for subsequent collaboration between agents. The shared information space is implemented using a MongoDB database, and its data schema includes fields for task ID, generating agent, content version, content data, and status identifier, ensuring data consistency and traceability during system operation.

[0022] Subsequently, patent-related knowledge data was acquired using big data networks. Natural language processing techniques were employed for word segmentation, entity recognition, and relation extraction to construct a patent rule knowledge graph stored in RDF triples. Based on this knowledge graph, a decision processing module was configured using the Drools rule engine to form an agent agent with professional legal knowledge processing capabilities. Simultaneously, technical documents were extracted from the patent-related data. The TF-IDF algorithm was used to calculate term weights, and key technical terms were filtered by setting thresholds. The TextRank algorithm was used to construct a term co-occurrence network based on a preset window, and the Louvain community discovery algorithm was combined to identify technical concept clusters. Finally, a technical knowledge base containing technical term nodes, technical effect nodes, and technical problem nodes was established. Based on the constructed technical knowledge base, a BERT pre-trained language model was deployed. Through continuous pre-training and fine-tuning, an expert agent was constructed, enabling it to deeply understand technical semantics. Finally, through a service registration and discovery mechanism and a message queue communication protocol, the collaborative agent, agent agent, and expert agent were integrated to construct a fully functional multi-agent writing assistance system.

[0023] Furthermore, in a preferred embodiment of the present invention, the establishment of a multi-agent writing assistance system composed of a collaborative intelligent agent, an agent intelligent agent, and an expert intelligent agent, and the configuration of a dedicated knowledge base and decision processing module for each intelligent agent, and the construction of a collaborative processing architecture consisting of a task scheduler, an arbitrator, and a shared information space, further includes: The workflow of the task scheduler includes: receiving a user request, parsing the technical solution text, generating an initial directed acyclic graph of tasks based on a preset template; using the Kahn topological sorting algorithm to determine the task execution sequence, asynchronously distributing task instructions through a message queue, and ensuring the reliability of task execution through heartbeat detection and timeout retry mechanisms. The arbitrator's conflict resolution mechanism includes: when a content conflict is detected, a multi-round debate process is initiated. The first round adopts a rapid ruling based on rule priority. If no consensus is reached, a voting mechanism based on confidence weighting is initiated. The debate log records the complete argument tree structure, including claim nodes, evidence nodes, and attack edges. The final ruling result is digitally signed to ensure immutability and is synchronized to each agent in real time. The data management process of the shared information space includes: adopting a version control mechanism based on content hashing and generating content fingerprints using the SHA-256 algorithm; setting up a full-text search interface with combined query modes based on time range, agent type, and task status to ensure complete data traceability and efficient access.

[0024] It should be noted that the task scheduler begins by receiving user requests, then parses the technical solution text, and generates an initial directed acyclic graph of tasks based on a pre-set patent disclosure template. This process uses the Kahn topological sorting algorithm to analyze the dependencies between task nodes, thereby determining the optimal task execution sequence. Task instructions are asynchronously dispatched through a message queue, achieving decoupling and efficient communication between system components. To ensure the reliability of the execution process, a heartbeat detection mechanism is introduced to monitor the running status of each agent in real time, and an outage retry mechanism automatically handles abnormal situations during task execution.

[0025] When the system detects content conflicts, the arbitrator initiates a multi-round debate process for coordination. The first round of debate employs a rapid adjudication mechanism based on rule priority, making decisions according to preset agent authority levels and rule weights. If the parties fail to reach a consensus, a confidence-weighted voting mechanism is activated, comprehensively considering the professional domain weights and historical performance scores of each agent. The entire debate process is fully recorded through an argument tree structure, including argument nodes, evidence nodes, and attack edges, ensuring the transparency and traceability of the reasoning process. The final adjudication result uses digital signature technology to ensure its immutability and is synchronized in real time to all relevant agents through a message broadcast mechanism.

[0026] The shared information space employs a content hash-based version control mechanism to manage data evolution, using the SHA-256 algorithm to generate a unique digital fingerprint for each content update. By establishing a chain of hash pointers between content versions, complete version history tracing of the data is achieved. To support efficient data retrieval, a full-text search interface based on multi-dimensional combined queries is designed, allowing users to flexibly query based on multiple dimensions such as time range, agent type, and task status.

[0027] Furthermore, in a preferred embodiment of the present invention, the original technical solution description received from the user is input into the multi-agent writing assistance system, which invokes collaborative agents to perform parallel semantic parsing and feature extraction to construct a structured intermediate representation containing the technical problem, technical solution, and technical effect, specifically including: The system obtains the original technical solution description input by the user and inputs it into the multi-agent writing assistance system. The collaborative agent uses a hybrid word segmentation algorithm that combines a patent domain dictionary to perform lexical-level segmentation on the original text, and identifies and replaces sensitive information and proper nouns in the text. The system constructs a lexical dependency tree through a dependency parser and marks part-of-speech tags and syntactic roles to generate a standardized text sequence. The standardized text sequence is distributed in parallel to the agent agent and the expert agent through the RabbitMQ message middleware; the agent agent uses the patent rule knowledge graph to generate a multi-dimensional rule distributed representation through the GraphSAGE graph neural network encoder, adopts a bidirectional LSTM-CRF architecture with 4 hidden layers and combines a multi-head self-attention mechanism to identify entities and limiting phrases in the text, and outputs patent rule feature vectors. Meanwhile, the expert agent captures the contextual information of the technical text through a multi-layer Transformer encoder, extracts technical entities using a pointer network-based BIO annotation sequence entity recognition mechanism, constructs the predicate-argument structure of the operation process through LTP dependency parsing and semantic role annotation, and finally obtains a multi-dimensional technical feature vector. The collaborative intelligent agent performs multi-stage feature alignment and fusion based on the patent rule feature vector and multi-dimensional technical feature vector, performs scale unification processing on the two types of feature vectors based on maximum and minimum normalization, and constructs a feature association matrix by calculating the cosine similarity between feature vectors. The feature association matrix is ​​regarded as the adjacency matrix of a weighted undirected graph. The corresponding degree matrix and Laplacian matrix are calculated. Then, the Laplacian matrix is ​​decomposed into features. The eigenvectors corresponding to the first k smallest eigenvalues ​​are selected to form a new feature space. In the new feature space, the K-means clustering algorithm is used to divide the original features into k different semantic clusters. For each feature cluster, the mean cosine similarity of the feature vectors within the cluster is calculated, and highly overlapping feature groups are identified by using a similarity threshold. Among the identified overlapping feature groups, the feature vector closest to the cluster center is retained as a representative feature, while other redundant feature vectors are removed to generate a filtered feature cluster. A multi-head attention mechanism is introduced to represent the query vector by patent rule feature vectors and key-value multi-dimensional technology feature vectors. Attention weights are calculated and feature fusion is performed to finally output multimodal fused features. Based on the multimodal fusion features, a syntactic dependency tree is constructed. A causal relationship graph with technical problems, technical means, and technical effects as causal paths is constructed through the syntactic dependency tree. The betweenness centrality and eigenvector centrality of each node in the causal relationship graph are used as node importance scores to construct a state space. A near-end policy optimization algorithm is used to perform semantic parsing optimization to output an optimized semantic parse tree. The optimized semantic parse tree is then mapped to a structured intermediate representation containing technical problems, technical solutions, and technical effects.

[0028] It should be noted that after receiving the original technical solution description input by the user, the collaborative agent performs deep preprocessing on the text. A hybrid word segmentation algorithm combining a patent domain dictionary is used to complete lexical-level segmentation, while identifying and replacing sensitive information and proper nouns in the text to protect trade secrets. Subsequently, a dependency parser is used to construct a lexical dependency tree, labeling it with part-of-speech tags and syntactic roles, forming a standardized text sequence as the basis for subsequent processing. Parallel distribution is achieved based on the preprocessed text sequence through the RabbitMQ message middleware. The agent agent, based on the patent rule knowledge graph, uses a graph neural network encoder to generate a multi-dimensional distributed representation of legal concepts. Through a bidirectional LSTM-CRF architecture with four hidden layers combined with a multi-head self-attention mechanism, it identifies legal entities and functional qualifying phrases in the text, outputting a hierarchical patent rule feature vector. At the same time, the expert agent captures deep contextual information of the technical text through a multi-layer Transformer encoder, extracts technical entities using a pointer network-based BIO-labeled sequence entity recognition mechanism, and constructs a predicate-argument structure of the operation flow with the help of LTP dependency parsing and semantic role labeling technology, finally outputting a multi-dimensional technical feature vector containing rich technical semantics.

[0029] After obtaining the two types of feature vectors, the collaborative agent initiates a multi-stage feature alignment and fusion process. First, the patent rule feature vector and the technical feature vector undergo scale unification processing based on min-max normalization to eliminate dimensional differences. A feature association matrix is ​​constructed by calculating the cosine similarity between feature vectors, characterizing the semantic association strength between legal and technical features. Then, the feature association matrix is ​​treated as an adjacency matrix of a weighted undirected graph, and the corresponding degree matrix and Laplacian matrix are calculated. By performing eigenvalue decomposition on the Laplacian matrix, the feature vectors corresponding to the k smallest eigenvalues ​​are selected to construct a dimensionality-reduced feature space. In the new feature space, the K-means clustering algorithm is used to divide the original features into k different semantic clusters. The mean cosine similarity of the feature vectors within each cluster is calculated, and highly overlapping feature groups are identified by setting a similarity threshold. Among the identified overlapping feature groups, the feature vector closest to the cluster center is retained as a representative feature to eliminate redundant information. Finally, a multi-head attention mechanism is introduced, using the patent rule feature vector as the query vector and the technical feature vector as the key vector. Deep interaction of features is achieved through dynamic allocation of attention weights, outputting multimodal fused features. A syntactic dependency tree is constructed based on multimodal fusion features, and a causal relationship graph is further generated. This graph clearly presents the causal relationship paths between technical problems, technical means, and technical effects. By calculating the betweenness centrality and eigenvector centrality of each node in the causal relationship graph, node importance scores are obtained, and a state space is constructed. The semantic parsing process is continuously optimized using a proximal policy optimization algorithm, outputting a well-structured optimized semantic parse tree. Finally, this parse tree is mapped to a structured intermediate representation containing technical problems, technical solutions, and technical effects, laying a solid foundation for the automatic generation of subsequent patent disclosure documents. The structured intermediate representation consists of hierarchical data objects that follow a specific pattern in the user's original technical description of the solution, clearly and systematically extracting and defining the core information in the original text.

[0030] Furthermore, in a preferred embodiment of the present invention, the step of generating a disclosure document writing task sequence based on the structured intermediate representation using a dynamic task decomposition algorithm, achieving task allocation and coordination among agents through the Contract Network protocol, and completing content generation and cross-validation based on a multi-round debate mechanism to obtain a draft disclosure document specifically includes: The collaborative intelligent agent performs dynamic task decomposition based on the structured intermediate representation, extracts the topological ordering relationship between each technical node, uses the critical path analysis method to determine the key sequence of task execution, and maps the entities in the structured intermediate representation and the relationships between entities to writing task nodes based on the chapter structure writing rules of the disclosure document, generating a directed acyclic graph of tasks with execution order. Based on the task directed acyclic graph, a contract network is used to conduct distributed task negotiation. The collaborative agent acts as the manager and encapsulates each subtask in the task directed acyclic graph into a task announcement containing task specifications, quality indicators and time constraints. The announcement is then broadcast to all registered agents through a message middleware. After all the agents receive the bidding information, they calculate the semantic matching degree between their own capability attributes and task requirements, combine real-time load status and historical performance indicators, and use multi-attribute utility theory to calculate the comprehensive bidding value. The collaborative agents conduct multi-objective optimization evaluation of all bids after the bidding deadline, and determine the optimal task-agent allocation mapping table based on the multi-objective optimization algorithm. Tasks are dispatched through an optimal task-agent allocation mapping table. The task-bearing agent generates an initial content draft based on its professional knowledge base. The agent agent is responsible for generating the rule content, and the expert agent is responsible for generating the technical content. The initial content draft is generated and cross-validated, with the technical content being validated by the expert agent and the rule content being validated by the agent agent. When content disputes are detected, the verification agent constructs a questioning framework that includes counterexamples and logical rebuttals, and initiates a debate challenge to the corresponding task agent through a message queue. After obtaining a consensus content draft through repeated debates, the content of the next task node is generated. The coordinating agent manages task execution based on the dependency relationships in the task dependency network graph. After all subtasks have completed debate verification, the collaborative agent synthesizes the final draft content of all subtasks based on the task execution order of the directed acyclic graph of tasks, and finally generates the draft disclosure document.

[0031] It should be noted that after obtaining the structured intermediate representation, the collaborative agent initiates a dynamic task decomposition process based on it. First, it analyzes the topological relationships between technical nodes and determines the critical sequence of task execution through critical path analysis. Following the chapter structure specifications of the patent disclosure, the technical entities and their semantic relationships in the structured intermediate representation are mapped to specific writing task nodes, including task units such as background technology description, invention content description, construction of specific implementation methods, and claim drafting. These task units, according to their inherent data dependencies, constitute a directed acyclic graph of tasks with a clearly defined execution order.

[0032] Based on the generated directed acyclic graph of tasks, a contract network protocol is used for distributed task negotiation. The collaborative agent, acting as the manager, encapsulates each subtask into a standardized task announcement containing task specifications, quality indicators, and time constraints, broadcasting it to all registered agents via a message broker. Upon receiving the bidding information, each agent calculates the semantic matching degree between its own capability attributes and task requirements, combining real-time load status and historical performance indicators, and uses multi-attribute utility theory to calculate the comprehensive bidding value. After the bidding deadline, the collaborative agent performs multi-objective optimization evaluation on all bids, determining the optimal task-agent allocation scheme based on the Pareto optimality principle, forming a complete task allocation mapping table. Content generation and verification processes are then performed based on the task allocation mapping table. The task-bearing agent generates an initial draft content based on its professional knowledge base, with the agent agent responsible for drafting rules such as claim layout and legal terminology, and the expert agent responsible for technical details such as technical solution descriptions and implementation construction. The generated draft content needs to undergo cross-validation. Technical content is verified for accuracy by other expert AI agents, and legal content is checked for compliance by other agent AI agents to ensure the professionalism and standardization of the content.

[0033] When content disputes are detected, a multi-round debate mechanism based on argumentation theory is initiated. The verification agent constructs a questioning framework including counterexamples and logical refutations, and initiates a debate challenge to the task-taking agent via a message queue. During the debate, both parties gradually eliminate cognitive differences through multiple rounds of assertion-refutation-correction dialogue. The collaborative agent monitors the debate progress in real time and intervenes when necessary according to pre-set arbitration rules. The content that reaches a consensus through the debate is revised and improved by the task-taking agent, and the new version is submitted to the version management system. After completing the debate verification of all sub-tasks, the collaborative agent systematically synthesizes the final draft content generated by each sub-task based on the execution order of the task's directed acyclic graph. During the synthesis process, ontology-based terminology unification, citation network-based cross-reference verification, and style rule-based format standardization are performed, ultimately generating a complete and compliant disclosure document draft.

[0034] Furthermore, in a preferred embodiment of the present invention, the step of performing multi-dimensional consistency checks on the generated draft disclosure document, coordinating ambiguities using a predefined conflict resolution strategy, eliminating logical conflicts and expression defects detected through an iterative revision process, and generating an optimized draft disclosure document for submission specifically includes: After generating the draft disclosure document, the collaborative intelligent agent constructs a consistency detection framework based on a multi-level graph structure and initializes the framework. The consistency detection framework includes four detection dimensions: technical logic dimension, writing compliance dimension, terminology consistency dimension, and graphic description matching dimension. The draft disclosure document is input into the initialized consistency detection framework for multi-dimensional parallel detection. The technical logic dimension detection is performed by an expert intelligent agent using a technical reasoning mechanism based on SWRL rules, and a technical logic detection report containing inconsistent node pairs and their confidence levels is generated. The agent performs compliance checks on the drafting process, parsing the draft disclosure document into a drafting element graph structure and using a violation pattern matching algorithm based on subgraph isomorphism to generate a compliance check report. The collaborative agent simultaneously performs terminology consistency and image-text matching checks, constructing a full-text terminology distribution vector space, calculating a terminology context similarity matrix, and combining it with an illustration keypoint matching algorithm based on SIFT feature extraction to generate a comprehensive check report that includes terminology consistency and image-text matching. After completing multi-dimensional parallel detection, the generated detection reports are integrated and a conflict evidence framework is constructed based on DS evidence theory. The conflict evidence framework is used to assign probability values ​​to each detected problem instance and calculate the trust function and likelihood function. Finally, a structured conflict detection report containing problem type, location coordinates, severity level, technical element correlation degree and resolution priority is generated.

[0035] It should be noted that after generating the draft disclosure document, the collaborative intelligent agent constructs and initializes a consistency detection framework based on a multi-level graph structure. This framework includes four core detection dimensions: the technical logic dimension focuses on analyzing the inherent causal relationships and logical consistency of the technical solution; the drafting compliance dimension ensures that the document structure complies with the requirements of patent law norms and examination guidelines; the terminology consistency dimension checks the accuracy and consistency of terminology used throughout the document; and the figure-text matching dimension verifies the consistency between the text description and the figure representation.

[0036] The draft disclosure document is input into the initialized detection framework, initiating a multi-dimensional parallel detection process. In the technical logic detection, the expert agent employs a SWRL-based technical reasoning mechanism, using semantic web rule language to logically verify the causal chains in the technical solution, identifying contradictory or broken technical node pairs, and calculating the confidence level of each issue based on rule matching, generating a technical logic detection report. Simultaneously, the agent agent performs drafting compliance detection, parsing the disclosure document content into a hierarchical drafting element graph, and using a subgraph isomorphism-based algorithm to match predefined violation patterns, accurately identifying various issues that do not conform to patent drafting specifications. The collaborative agent simultaneously conducts terminology consistency and image-text matching detection. By constructing a full-text terminology distribution vector space and calculating the semantic similarity matrix of terms in context, the system can identify areas of inconsistent terminology usage. Simultaneously, a SIFT-based feature extraction algorithm is used to analyze key points in illustrations, performing multimodal matching degree calculations with corresponding text descriptions to ensure high consistency between text and images, ultimately generating a comprehensive detection report.

[0037] After completing parallel detection across all dimensions, a unified conflict evidence framework is constructed based on Dempster's evidence theory. This framework transforms issue instances from each dimension's detection report into independent sources of evidence. It calculates the joint probability distribution for each issue using Dempster's combination rules, assigns a basic probability value to each detected issue instance, and calculates the corresponding trust and likelihood functions, ultimately generating a structured conflict detection report. This report records the type attribute of each issue, its specific location coordinates in the document, severity level, the correlation of involved technical elements, and the system's recommended resolution priority, providing precise data support for subsequent revisions and optimizations.

[0038] Furthermore, in a preferred embodiment of the present invention, the step of performing multi-dimensional consistency checks on the generated draft disclosure document, coordinating ambiguities using a predefined conflict resolution strategy, eliminating logical conflicts and expression defects detected through an iterative revision process, and generating an optimized draft disclosure document for submission further includes: The collaborative intelligent agent uses the TOPSIS multi-attribute decision algorithm to rank conflict problems based on the structured conflict detection report. It sets an objective function that includes the severity of the conflict, the scope of its impact, and the complexity of its processing, determines the priority sequence of problem processing, and obtains a ranked conflict problem sequence. The collaborative intelligent agent then obtains conflict resolution strategies in the shared information space through the ranked conflict problem sequence. For technical feature description conflicts, semantic similarity between different expressions is calculated, and word vector models are used to generate modification schemes for description conflicts; for drafting compliance conflicts, the agent's intelligent agent is invoked to generate drafting modification schemes based on the patent rule knowledge graph; for logical chain conflicts, the expert intelligent agent is invoked to monitor and generate logical connection modification schemes in the technical field knowledge base; finally, a set of conflict resolution strategies is generated. The conflict resolution priority is extracted by sorting the conflict problem sequence, and the draft disclosure document is iteratively revised in combination with the conflict resolution strategy set. In each iteration, a consistency check is performed again based on the generated new draft disclosure document. If there are conflict problems that do not achieve the expected resolution effect, the corresponding resolution strategy is marked as an inefficient strategy and a new resolution strategy is obtained. Through an iterative process of repeated detection, revision, and verification until the resolution of all conflicts reaches a preset quality threshold, the final optimized draft disclosure document is output and pushed out.

[0039] It should be noted that the collaborative intelligent agent, based on the structured conflict detection report, uses the TOPSIS multi-attribute decision algorithm to intelligently rank the detected conflict issues. This algorithm comprehensively considers multiple objective functions such as conflict severity, impact scope, and processing complexity, and generates a priority-ranked sequence of conflict issues by calculating the relative closeness of each issue to the ideal solution. Based on this sequence, corresponding solutions are retrieved from a strategy library in the shared information space. For conflicts with inconsistent technical feature descriptions, semantic similarity is calculated using a word vector model to generate a terminology unification scheme. For compliance drafting conflicts, the agent intelligent agent provides normative modification suggestions based on a patent rule knowledge graph. For conflicts with broken logical chains, the expert intelligent agent extracts relevant knowledge from the technical field knowledge base to generate logical connection schemes, ultimately forming a set of conflict resolution strategies. Subsequently, the draft disclosure document is iteratively revised multiple times according to the priority order of the conflict issues and the set of resolution strategies. After each revision, a consistency check is re-executed to evaluate the actual effectiveness of the resolution strategies. For conflict issues that do not achieve the expected resolution effect, the corresponding resolution strategy is marked as an inefficient strategy, and an alternative strategy retrieval mechanism is activated to obtain new solutions from the strategy library. Through this iterative process of detection-revision-verification, the solution is continuously optimized until the resolution of all conflicts reaches the preset quality threshold. The final output is a fully optimized draft of the disclosure document, which, along with the complete revision history, is pushed to the user via a push notification interface, ensuring the reliability of the deliverables and the traceability of the process.

[0040] Figure 2 A flowchart of a second method for assisting in drafting a patent disclosure document based on multi-agent collaboration, provided as an embodiment of the present invention; like Figure 2 As shown, the present invention provides a second method flowchart for a patent disclosure document drafting assistance method based on multi-agent collaboration, comprising: S202: Extract the user's original technical solution description from the shared information space and obtain the optimized disclosure draft. Perform word segmentation and dependency parsing on the original technical solution description based on the patent field dictionary to identify the core technical problem statement and innovation point description. At the same time, perform structured parsing on the optimized disclosure draft to extract key technical terms, technical problem focus and technical effect features to form an initial user interest feature set. S204, the initial set of user interest features is input into the expert agent, the technical knowledge base is used to perform semantic expansion and association analysis on each user interest feature, the semantic similarity between technical terms is used to identify potential related technical concepts, the LDA topic analysis is used to extract the technical field distribution corresponding to the original technical solution, and finally a multi-dimensional interest feature vector containing core interest features, related technical concepts and technical field preferences is constructed. S206, Based on the multi-dimensional interest feature vector, the user interest features are integrated according to four dimensions: technical field, technical problem, technical means and technical effect, and a user interest profile is constructed and a weight value is set for each dimension. The weight value is determined by the frequency of the corresponding interest feature in the original technical solution and the optimization disclosure document. S208, the user interest profile is converted into a retrieval query vector, a big data network is introduced to perform interest document retrieval through semantic similarity calculation, and the retrieval results are sorted by calculating the semantic relevance between the query vector and the document content, and interest documents within a preset range are selected to generate an interest retrieval result set. S210, the interest retrieval result set is imported into the expert intelligent agent for analysis. The technical description features corresponding to each interest document are extracted by key phrase extraction technology to construct the interest summary of each interest document and compare it with the original technical solution description to obtain the difference analysis results. Finally, an interest technology content report is generated and pushed to the target user.

[0041] It should be noted that, firstly, the original technical solution description and the optimized draft disclosure document of the user are extracted from the shared information space. Through word segmentation and dependency parsing based on a patent domain dictionary, the core technical problem statements and innovative points in the original solution are identified. Simultaneously, the optimized draft disclosure document is structurally parsed to extract key technical terms, technical problem focuses, and technical effect features. These features together constitute the initial user interest feature set, laying the data foundation for subsequent in-depth analysis. Subsequently, the initial user interest feature set is input into an expert agent for further processing. The expert agent utilizes its technical field knowledge base to perform semantic expansion and correlation analysis on each interest feature. By calculating the semantic similarity between technical terms, potential related technical concepts are identified. At the same time, LDA topic modeling is used to analyze the domain distribution characteristics of the technical problem. Finally, a multi-dimensional interest feature vector containing core interest features, related technical concepts, and technical field preferences is constructed, comprehensively characterizing the user's technical concerns.

[0042] Based on multi-dimensional interest feature vectors, a structured user interest profile is constructed according to four dimensions: technical field, technical problem, technical means, and technical effect. The feature weights under each dimension are dynamically determined by statistically analyzing the frequency of occurrence of the corresponding features in the original technical solution and the optimized disclosure document, ensuring that the weight allocation accurately reflects the user's true technical preferences. This interest profile not only includes the user's explicitly expressed technical needs but also covers potential interest directions discovered through semantic analysis. Subsequently, the user interest profile is converted into a retrieval query vector, and interest document retrieval is performed through a big data network. During the retrieval process, semantic similarity calculation is used to match the query vector with the document content, and the retrieval results are sorted according to semantic relevance, selecting documents highly relevant to the user's interests to form an interest retrieval result set. This result set includes patent documents as well as academic papers and technical reports from multiple sources. Finally, an expert intelligent agent performs in-depth analysis of the interest retrieval result set, extracting the technical features of each document through key phrase extraction technology to construct standardized interest summaries. These summaries are compared with the original technical solutions to identify the technological development path and innovation opportunities, ultimately generating an interest technology content report containing technology trend analysis and innovation suggestions, which is then pushed to the user via an interface, providing professional reference for subsequent technological innovation.

[0043] Furthermore, the patent disclosure writing assistance method based on multi-agent collaboration provided by this invention also includes the following steps: When an agent raises questions or requests for supplements to the draft content, the agent obtains the question feedback text information, inputs the question feedback text information into the collaborative intelligent agent, identifies the technical features and core demands involved in the question text, and constructs a structured question representation that includes elements such as question type, technical field, problem focus, and expected supplementary content. Content demand prediction is performed based on the structured question representation. The feedback content features of the agent in the feedback text are extracted through the structured question representation and input into the expert agent to perform content demand prediction through the built-in graph neural network, thereby generating content demand prediction information. Using the content demand prediction information as a retrieval vector, a retrieval content set is obtained by searching the technical knowledge base associated with the expert agent. At the same time, the content demand information is input into the agent agent to obtain the drafting rule requirements involved in the question based on the patent rule knowledge base, and a question analysis report containing technical supplementary requirements, legal normative explanations and evidence suggestions is generated. Based on the question analysis report, the scope of technical details and example types that need to be supplemented are generated, and supplementary content templates are generated according to the writing rules to form a feedback content planning scheme. Based on the feedback content planning scheme, user-oriented feedback suggestion text is generated by combining the expression characteristics of the user's original technical solution with a pre-trained text generator to assist the target user in understanding the question.

[0044] It should be noted that during the patent drafting process, the questions raised by patent attorneys based on their professional experience often involve complex technical details and legal requirements. Inventors, due to limitations in their professional background, often struggle to accurately understand the attorney's professional demands, leading to multiple rounds of communication before reaching a consensus. This communication barrier not only prolongs the patent preparation cycle but can also affect the final patent quality due to distorted information transmission. This solution, by constructing an intelligent question feedback and processing mechanism, effectively bridges the gap between the attorney's professional requirements and the inventor's comprehension capabilities. Its core significance lies in establishing a bridge from professional questions to user-friendly guidance, transforming professional legal terminology and technical requirements into clear, actionable guidance for inventors through multi-agent collaborative analysis. This transformation not only resolves the professional gap but, more importantly, establishes a standardized question-handling process, ensuring that every question is handled systematically and in a standardized manner. Furthermore, it significantly improves communication efficiency by intelligently analyzing the agent's professional questions and generating targeted guidance, optimizing traditional multi-round communication into one-time precise guidance. Secondly, it ensures the accurate transmission of professional requirements by using a graph neural network-based demand prediction and dual knowledge base verification mechanism to ensure that technical requirements and legal norms are transmitted completely and accurately. Finally, it transforms professional requirements into user-friendly expressions through personalized text generation technology, enabling inventors to quickly and accurately complete content supplementation and providing strong support for patent communication and drafting.

[0045] Figure 3 A patent disclosure document drafting assistance system 3 based on multi-agent collaboration is provided as an embodiment of the present invention. The system includes: a user interaction module, a collaborative agent module, an agent agent module, an expert agent module, a data storage module, and an optimization control module. The user interaction module is used to handle user input and output interaction, including receiving the initial technical solution description, displaying the generated results, and collecting user feedback. It is implemented through a graphical user interface or API interface, supports natural language input, and updates the progress status in real time. The collaborative intelligent agent module, as the central coordinator of the system, is responsible for parsing user input, assigning tasks and driving iterative optimization. It uses natural language processing technology to extract key information, calls other intelligent agent modules, and compares the output content to generate a difference report, so as to achieve multi-round interaction and feedback integration. The agent intelligent agent module is used for the standardized writing of patent disclosure documents. It has a built-in patent rule knowledge graph and template engine, and generates a text structure that conforms to the patent format through rule matching and grammatical analysis. The expert agent module is responsible for improving and supplementing the technical solution, integrating machine learning models and technical knowledge bases, identifying missing points in the technical solution through semantic analysis, and recommending technical directions of interest. The data storage module is used to manage the system's persistent data, including rule base, knowledge base, user history records and temporary cache. It stores and retrieves data through a database system, supports efficient querying and updating, and ensures the consistency of data sharing among modules. The optimization control module is used to invoke innovative optimization suggestions, requiring the agent to continue optimizing the technical solution.

[0046] A third aspect of the present invention provides a computer-readable storage medium comprising a method program for assisting in drafting a patent disclosure document based on multi-agent collaboration. When the method program is executed by a processor, it implements the steps of the method program for assisting in drafting a patent disclosure document based on multi-agent collaboration as described in any of the preceding claims.

[0047] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.

[0048] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units. They may be located in one place or distributed across multiple network units. Some or all of the units may be selected to achieve the purpose of this embodiment according to actual needs.

[0049] In addition, in the various embodiments of the present invention, each functional unit can be integrated into one processing unit, or each unit can be a separate unit, 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.

[0050] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0051] Alternatively, if the integrated units of this invention are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROM, RAM, magnetic disks, or optical disks.

[0052] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for assisting in drafting patent disclosure documents based on multi-agent collaboration, characterized in that, include: Establish a multi-agent writing assistance system consisting of collaborative agents, agent agents, and expert agents, and configure dedicated knowledge bases and decision processing modules for each agent. Construct a collaborative processing architecture with a task scheduler, arbitrator, and shared information space. The system receives the original technical solution description input by the user and inputs it into the multi-agent writing assistance system to call on collaborative multi-agents to perform parallel semantic parsing and feature extraction, and constructs a structured intermediate representation that includes technical problems, technical solutions and technical effects. Based on the structured intermediate representation, a dynamic task decomposition algorithm is used to generate a task sequence for writing the disclosure document. The task allocation and coordination among agents are realized through the contract network protocol. The content generation and cross-validation are completed based on a multi-round debate mechanism to obtain the draft disclosure document. The generated draft disclosure document undergoes multi-dimensional consistency checks, and predefined conflict resolution strategies are used to coordinate ambiguities. Logical conflicts and expression defects are eliminated through an iterative revision process, and an optimized draft disclosure document is generated and pushed out. Based on the original technical solution description, a user interest profile is generated by combining the optimized disclosure draft, and an interest retrieval is performed based on the user interest profile to generate an interest-based technical content report.

2. The method for assisting in drafting a patent disclosure document based on multi-agent collaboration as described in claim 1, characterized in that, The establishment of a multi-agent writing assistance system consisting of collaborative agents, agent agents, and expert agents, and the configuration of dedicated knowledge bases and decision processing modules for each agent, along with the construction of a collaborative processing architecture comprising a task scheduler, an arbitrator, and a shared information space, specifically includes: Deploy independent microservice instances in a distributed computing environment, integrate a workflow engine based on Apache Airflow as a task scheduler, the workflow engine is pre-configured with several standard chapter templates of patent disclosure documents, and the execution flow of the directed acyclic graph is constructed between the chapters through data dependencies, and a collaborative processing architecture of arbitrator, shared information space and task scheduler is constructed to obtain a collaborative intelligent agent; The arbitrator is implemented using a rule engine based on Drools. The loaded conflict resolution rule base includes a set of rules for agent priority, a set of rules for content consistency, and a set of rules for semantic conflict detection. The shared information space is implemented using a MongoDB database. The data pattern of the shared information space includes fields for task ID, generating agent, content version, content data, and status identifier. Patent-related knowledge data is obtained by using big data networks, patent drafting rule data is extracted, and natural language processing technology is used for word segmentation, entity recognition and relation extraction. A patent rule knowledge graph stored in the form of RDF triples is constructed. The decision processing module is configured through the Drools rule engine using the patent rule knowledge graph, and finally the agent intelligent agent is obtained. Technical documents are extracted from the patent-related knowledge data, term weights are calculated using the TF-IDF algorithm, key technical terms are filtered by setting thresholds, a term co-occurrence network is constructed using the TextRank algorithm based on a pre-set co-occurrence window, and clusters of technical concepts are identified using the Louvain community discovery algorithm. A technical knowledge base containing technical term nodes, technical effect nodes, and technical problem nodes is established. Deploy a BERT pre-trained language model and continuously pre-train and fine-tune it using technical documents in the knowledge base of the aforementioned technical field to construct an expert agent; by integrating the collaborative agent, agent agent, and expert agent, and setting a service registration and discovery mechanism and a message queue communication protocol, construct a multi-agent writing assistance system.

3. The method for assisting in drafting a patent disclosure document based on multi-agent collaboration as described in claim 1, characterized in that, The establishment of a multi-agent writing assistance system consisting of collaborative agents, agent agents, and expert agents, and the configuration of dedicated knowledge bases and decision processing modules for each agent, along with the construction of a collaborative processing architecture comprising a task scheduler, an arbitrator, and a shared information space, further includes: The workflow of the task scheduler includes: receiving a user request, parsing the technical solution text, generating an initial directed acyclic graph of tasks based on a preset template; using the Kahn topological sorting algorithm to determine the task execution sequence, asynchronously distributing task instructions through a message queue, and ensuring the reliability of task execution through heartbeat detection and timeout retry mechanisms. The arbitrator's conflict resolution mechanism includes: when a content conflict is detected, a multi-round debate process is initiated. The first round adopts a rapid ruling based on rule priority. If no consensus is reached, a voting mechanism based on confidence weighting is initiated. The debate log records the complete argument tree structure, including claim nodes, evidence nodes, and attack edges. The final ruling result is digitally signed to ensure immutability and is synchronized to each agent in real time. The data management process of the shared information space includes: adopting a version control mechanism based on content hashing and generating content fingerprints using the SHA-256 algorithm; setting up a full-text search interface with combined query modes based on time range, agent type, and task status to ensure complete data traceability and efficient access.

4. The method for assisting in drafting a patent disclosure document based on multi-agent collaboration as described in claim 1, characterized in that, The original technical solution description received from the user is input into the multi-agent writing assistance system, which calls on collaborative agents to perform parallel semantic parsing and feature extraction to construct a structured intermediate representation containing the technical problem, technical solution, and technical effect. Specifically, this includes: The system obtains the original technical solution description input by the user and inputs it into the multi-agent writing assistance system. The collaborative agent uses a hybrid word segmentation algorithm that combines a patent domain dictionary to perform lexical-level segmentation on the original text, and identifies and replaces sensitive information and proper nouns in the text. The system constructs a lexical dependency tree through a dependency parser and marks part-of-speech tags and syntactic roles to generate a standardized text sequence. The standardized text sequence is distributed in parallel to the agent agent and the expert agent through the RabbitMQ message middleware; the agent agent uses the patent rule knowledge graph to generate a multi-dimensional rule distributed representation through the GraphSAGE graph neural network encoder, adopts a bidirectional LSTM-CRF architecture with 4 hidden layers and combines a multi-head self-attention mechanism to identify entities and limiting phrases in the text, and outputs patent rule feature vectors. Meanwhile, the expert agent captures the contextual information of the technical text through a multi-layer Transformer encoder, extracts technical entities using a pointer network-based BIO annotation sequence entity recognition mechanism, constructs the predicate-argument structure of the operation process through LTP dependency parsing and semantic role annotation, and finally obtains a multi-dimensional technical feature vector. The collaborative intelligent agent performs multi-stage feature alignment and fusion based on the patent rule feature vector and multi-dimensional technical feature vector, performs scale unification processing on the two types of feature vectors based on maximum and minimum normalization, and constructs a feature association matrix by calculating the cosine similarity between feature vectors. The feature association matrix is ​​regarded as the adjacency matrix of a weighted undirected graph. The corresponding degree matrix and Laplacian matrix are calculated. Then, the Laplacian matrix is ​​decomposed into features. The eigenvectors corresponding to the first k smallest eigenvalues ​​are selected to form a new feature space. In the new feature space, the K-means clustering algorithm is used to divide the original features into k different semantic clusters. For each feature cluster, the mean cosine similarity of the feature vectors within the cluster is calculated, and highly overlapping feature groups are identified by using a similarity threshold. Among the identified overlapping feature groups, the feature vector closest to the cluster center is retained as a representative feature, while other redundant feature vectors are removed to generate a filtered feature cluster. A multi-head attention mechanism is introduced to represent the query vector by patent rule feature vectors and key-value multi-dimensional technology feature vectors. Attention weights are calculated and feature fusion is performed to finally output multimodal fused features. Based on the multimodal fusion features, a syntactic dependency tree is constructed. A causal relationship graph with technical problems, technical means, and technical effects as causal paths is constructed through the syntactic dependency tree. The betweenness centrality and eigenvector centrality of each node in the causal relationship graph are used as node importance scores to construct a state space. A near-end policy optimization algorithm is used to perform semantic parsing optimization to output an optimized semantic parse tree. The optimized semantic parse tree is then mapped to a structured intermediate representation containing technical problems, technical solutions, and technical effects.

5. The method for assisting in drafting a patent disclosure document based on multi-agent collaboration as described in claim 1, characterized in that, Based on the structured intermediate representation, a dynamic task decomposition algorithm is used to generate a task sequence for writing the disclosure document. The contract network protocol is used to achieve task allocation and coordination among agents. A multi-round debate mechanism is used to complete content generation and cross-validation, resulting in a draft disclosure document, which specifically includes: The collaborative intelligent agent performs dynamic task decomposition based on the structured intermediate representation, extracts the topological ordering relationship between each technical node, uses the critical path analysis method to determine the key sequence of task execution, and maps the entities in the structured intermediate representation and the relationships between entities to writing task nodes based on the chapter structure writing rules of the disclosure document, generating a directed acyclic graph of tasks with execution order. Based on the task directed acyclic graph, a contract network is used to conduct distributed task negotiation. The collaborative agent acts as the manager and encapsulates each subtask in the task directed acyclic graph into a task announcement containing task specifications, quality indicators and time constraints. The announcement is then broadcast to all registered agents through a message middleware. After all the agents receive the bidding information, they calculate the semantic matching degree between their own capability attributes and task requirements, combine real-time load status and historical performance indicators, and use multi-attribute utility theory to calculate the comprehensive bidding value. The collaborative agents conduct multi-objective optimization evaluation of all bids after the bidding deadline, and determine the optimal task-agent allocation mapping table based on the multi-objective optimization algorithm. Tasks are dispatched through an optimal task-agent allocation mapping table. The task-bearing agent generates an initial content draft based on its professional knowledge base. The agent agent is responsible for generating the rule content, and the expert agent is responsible for generating the technical content. The initial content draft is generated and cross-validated, with the technical content being validated by the expert agent and the rule content being validated by the agent agent. When content disputes are detected, the verification agent constructs a questioning framework that includes counterexamples and logical rebuttals, and initiates a debate challenge to the corresponding task agent through a message queue. After obtaining a consensus content draft through repeated debates, the content of the next task node is generated. The coordinating agent manages task execution based on the dependency relationships in the task dependency network graph. After all subtasks have completed debate verification, the collaborative agent synthesizes the final draft content of all subtasks based on the task execution order of the directed acyclic graph of tasks, and finally generates the draft disclosure document.

6. The method for assisting in drafting a patent disclosure document based on multi-agent collaboration as described in claim 1, characterized in that, The process involves performing multi-dimensional consistency checks on the generated draft disclosure document, using predefined conflict resolution strategies to coordinate ambiguities, and eliminating logical conflicts and expression defects through an iterative revision process to generate an optimized draft disclosure document for submission. Specifically, this includes: After generating the draft disclosure document, the collaborative intelligent agent constructs a consistency detection framework based on a multi-level graph structure and initializes the framework. The consistency detection framework includes four detection dimensions: technical logic dimension, writing compliance dimension, terminology consistency dimension, and graphic description matching dimension. The draft disclosure document is input into the initialized consistency detection framework for multi-dimensional parallel detection. The technical logic dimension detection is performed by an expert intelligent agent using a technical reasoning mechanism based on SWRL rules, and a technical logic detection report containing inconsistent node pairs and their confidence levels is generated. The agent performs compliance checks on the drafting process, parsing the draft disclosure document into a drafting element graph structure and using a violation pattern matching algorithm based on subgraph isomorphism to generate a compliance check report. The collaborative agent simultaneously performs terminology consistency and image-text matching checks, constructing a full-text terminology distribution vector space, calculating a terminology context similarity matrix, and combining it with an illustration keypoint matching algorithm based on SIFT feature extraction to generate a comprehensive check report that includes terminology consistency and image-text matching. After completing multi-dimensional parallel detection, the generated detection reports are integrated and a conflict evidence framework is constructed based on DS evidence theory. The conflict evidence framework is used to assign probability values ​​to each detected problem instance and calculate the trust function and likelihood function. Finally, a structured conflict detection report containing problem type, location coordinates, severity level, technical element correlation degree and resolution priority is generated.

7. The method for assisting in drafting a patent disclosure document based on multi-agent collaboration as described in claim 1, characterized in that, The process of performing multi-dimensional consistency checks on the generated draft disclosure document, coordinating ambiguities using predefined conflict resolution strategies, eliminating logical conflicts and expression defects through an iterative revision process, and generating an optimized draft disclosure document for submission also includes: The collaborative intelligent agent uses the TOPSIS multi-attribute decision algorithm to rank conflict problems based on the structured conflict detection report. It sets an objective function that includes the severity of the conflict, the scope of its impact, and the complexity of its processing, determines the priority sequence of problem processing, and obtains a ranked conflict problem sequence. The collaborative intelligent agent then obtains conflict resolution strategies in the shared information space through the ranked conflict problem sequence. For technical feature description conflicts, semantic similarity between different expressions is calculated, and word vector models are used to generate modification schemes for description conflicts; for drafting compliance conflicts, the agent's intelligent agent is invoked to generate drafting modification schemes based on the patent rule knowledge graph; for logical chain conflicts, the expert intelligent agent is invoked to monitor and generate logical connection modification schemes in the technical field knowledge base; finally, a set of conflict resolution strategies is generated. The conflict resolution priority is extracted by sorting the conflict problem sequence, and the draft disclosure document is iteratively revised in combination with the conflict resolution strategy set. In each iteration, a consistency check is performed again based on the generated new draft disclosure document. If there are conflict problems that do not achieve the expected resolution effect, the corresponding resolution strategy is marked as an inefficient strategy and a new resolution strategy is obtained. Through an iterative process of repeated detection, revision, and verification until the resolution of all conflicts reaches a preset quality threshold, the final optimized draft disclosure document is output and pushed out.

8. The method for assisting in drafting a patent disclosure document based on multi-agent collaboration as described in claim 1, characterized in that, The process of generating a user interest profile based on the original technical solution description and the optimized draft disclosure document, performing interest retrieval based on the user interest profile, and generating an interest-related technical content report specifically includes: The original technical solution description of the user is extracted from the shared information space and the optimized disclosure draft is obtained. The original technical solution description is processed by word segmentation and dependency parsing based on the patent field dictionary to identify the core technical problem statement and the description of innovation points. At the same time, the optimized disclosure draft is subjected to structured parsing to extract key technical terms, technical problem focus and technical effect features to form an initial set of user interest features. The initial set of user interest features is input into the expert agent. The technical knowledge base is used to perform semantic expansion and association analysis on each user interest feature. Potential related technical concepts are identified by the semantic similarity between technical terms. The technical field distribution corresponding to the original technical solution is extracted by LDA topic analysis. Finally, a multi-dimensional interest feature vector containing core interest features, related technical concepts and technical field preferences is constructed. Based on the multidimensional interest feature vector, user interest features are integrated according to four dimensions: technical field, technical problem, technical means and technical effect, and a user interest profile is constructed. A weight value is set for each dimension, wherein the weight value is determined by the frequency of the corresponding interest feature in the original technical solution and the optimization disclosure document. The user interest profile is converted into a retrieval query vector, and a big data network is introduced to retrieve interest documents through semantic similarity calculation. The retrieval results are sorted by calculating the semantic relevance between the query vector and the document content, and interest documents within a preset range are selected to generate an interest retrieval result set. The interest retrieval result set is imported into an expert intelligent agent for analysis. The technical description features corresponding to each interest document are extracted using key phrase extraction technology to construct an interest summary for each interest document. This summary is then compared with the original technical solution description to obtain the difference analysis results. Finally, an interest technology content report is generated and pushed to the target users.

9. A patent disclosure document drafting assistance system based on multi-agent collaboration, characterized in that, The system includes: a user interaction module, a collaborative intelligent agent module, an agent intelligent agent module, an expert intelligent agent module, a data storage module, and an optimization control module; The user interaction module is used to handle user input and output interaction, including receiving the initial technical solution description, displaying the generated results, and collecting user feedback. It is implemented through a graphical user interface or API interface, supports natural language input, and updates the progress status in real time. The collaborative intelligent agent module, as the central coordinator of the system, is responsible for parsing user input, assigning tasks and driving iterative optimization. It uses natural language processing technology to extract key information, calls other intelligent agent modules, and compares the output content to generate a difference report, so as to achieve multi-round interaction and feedback integration. The agent intelligent agent module is used for the standardized writing of patent disclosure documents. It has a built-in patent rule knowledge graph and template engine, and generates a text structure that conforms to the patent format through rule matching and grammatical analysis. The expert agent module is responsible for improving and supplementing the technical solution, integrating machine learning models and technical knowledge bases, identifying missing points in the technical solution through semantic analysis, and recommending technical directions of interest. The data storage module is used to manage the system's persistent data, including rule base, knowledge base, user history records and temporary cache. It stores and retrieves data through a database system, supports efficient querying and updating, and ensures the consistency of data sharing among modules. The optimization control module is used to invoke innovative optimization suggestions, requiring the agent to continue optimizing the technical solution.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a patent disclosure document drafting assistance method program based on multi-agent collaboration. When the program is executed by a processor, it implements the steps of the patent disclosure document drafting assistance method based on multi-agent collaboration as described in any one of claims 1 to 8.

Images