Patent text structuring processing and multi-hop semantic query method and system

Abstract

The application belongs to the technical field of data processing, and provides a patent text structured processing and multi-hop semantic query method and system. The method comprises the following steps: obtaining a patent file to be processed, performing structured cleaning and format regularization on the patent file to be processed, generating a standardized input text, and storing the key structured information obtained by the structured cleaning into a relational database; according to the standardized input text, an entity and relationship recognition are combined with clustering and global indexing to construct a patent knowledge graph; after receiving a query question of a user, the query intention type of the user is determined, and the relational database and / or the patent knowledge graph are called according to the query intention type of the user to perform multi-hop semantic query, and a semantic query result is obtained; according to the semantic query result, an answer to the query question is generated and fed back to the user. The scheme improves the accuracy, comprehensiveness and reliability of the patent query result, and optimizes the interactive experience and query efficiency of the patent retrieval.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a patented method and system for text structuring and multi-hop semantic querying. Background Technology

[0002] Currently, patent retrieval and analysis solutions for patent file management scenarios are mainly divided into two categories: field matching retrieval based on structured databases and patent information query based on retrieval enhancement. Both have significant technical defects in practical applications and cannot meet the needs of file management for patent retrieval accuracy, semantic understanding ability, and the ability to balance multi-hop logical reasoning and statistical analysis.

[0003] Specifically, field-matching retrieval methods based on structured databases rely on pre-structured fields such as application number, applicant, and classification number in the patent database. Retrieval is achieved through matching filtering conditions or SQL queries. This method can only achieve precise matching at the field level and cannot understand the technical semantics of long texts such as patent specifications and claims. It also cannot identify deeper information such as synonyms and technical feature associations. Furthermore, it is unable to handle multi-hop reasoning or complex technical relationship analysis across fields and documents. The results are highly dependent on the accuracy of the filtering conditions and are prone to missed or false detections, making it difficult to support advanced retrieval needs such as innovation analysis and trend insight.

[0004] The patent information query method based on retrieval enhancement mainly combines vector retrieval and large language models to achieve semantic-level query. Although it can understand the technical semantics in the patent text and support a certain degree of fuzzy query and cross-paragraph association, the retrieval process relies on vector recall and cannot establish a stable entity and relation structure. In multi-hop logical reasoning tasks with cross-paragraph and multi-entity association, it is easy to miss key links, resulting in poor stability of the generated results. When faced with the complex problem of combining statistical query and summary analysis, this method relies solely on model inference, which is prone to problems such as inaccurate numerical values, incomplete logic, or reasoning gaps.

[0005] It is evident that traditional patent search and analysis methods suffer from technical problems such as low processing accuracy, poor reliability, and lack of comprehensiveness and efficiency. Summary of the Invention

[0006] This invention provides a method and system for structuring patent text and performing multi-hop semantic queries, which addresses the shortcomings of traditional patent retrieval and analysis schemes, such as low processing accuracy, poor reliability, and lack of comprehensiveness and efficiency.

[0007] On the one hand, this invention provides a patented text structuring processing and multi-hop semantic query method, including: Obtain the patent document to be processed, perform structured cleaning and format regularization on the patent document to be processed, generate standardized input text, and store the key structured information obtained from the structured cleaning into a relational database; Based on the standardized input text, a patent knowledge graph is constructed by combining entity and relationship recognition with clustering and global indexing. After receiving the user's query question, the query intent type is determined, and based on the query intent type, the relational database and / or the patent knowledge graph are invoked to perform a multi-hop semantic query to obtain the semantic query result. Based on the semantic query results, the answer to the query question is generated and returned to the user.

[0008] According to the patent text structuring and multi-hop semantic query method provided by the present invention, the patent document to be processed is subjected to structuring cleaning and format regularization to generate standardized input text, including: The patent documents to be processed are initially cleaned and converted to a specific format to obtain a document in the set format. The document with the specified format is structurally segmented to identify multiple key blocks; Deep cleaning is performed on the content of each key block to obtain the cleaned content of each key block; The cleaned content of all key blocks is integrated and formatted to generate standardized input text.

[0009] According to the patent text structuring and multi-hop semantic query method provided by the present invention, the multiple key blocks include: bibliographic items, claims, and specification. Deep cleaning is performed on the content of each key block to obtain the cleaned content of each key block, including: Identify and clean the key fields in the bibliographic data, perform special rule-based processing on the multiple target contents of the key fields, and obtain a bibliographic structure dictionary, which is used as the cleaned content of the bibliographic data. Identify and extract the valid content of each claim in the claim book by segment according to the number, and perform format standardization and text cleaning on the valid content of each claim, output the claim list, and use it as the cleaned content of the claim book; Key sections of the specification are extracted, paragraphs are organized, and the invention content and specific embodiments in the organized key sections are assembled to generate an overview of the specification, which serves as the cleaned-up content of the specification.

[0010] According to the patent text structuring and multi-hop semantic query method provided by the present invention, based on the standardized input text, a patent knowledge graph is constructed by combining entity and relation recognition with clustering and global indexing, including: The standardized input text is subjected to text import and basic preprocessing to obtain preprocessed text; According to the inherent structure of the patent document, the preprocessed text is semantically segmented and multiple content units are constructed. Extract the patent technology entities from the multiple content units, identify the entity technology relationships between the patent technology entities, and construct an initial graph structure based on the patent technology entities and the entity technology relationships; By clustering and global indexing, the initial graph structure is post-processed to construct a patent knowledge graph.

[0011] According to the patent text structuring and multi-hop semantic query method provided by the present invention, the initial graph structure is post-processed through clustering and global indexing to obtain a patent knowledge graph, including: The initial graph structure is subjected to topic clustering to obtain semantically related local knowledge units and corresponding topic explanation texts; Based on the initial graph structure, the local knowledge units, and the topic explanation text, a patent knowledge graph is generated through global indexing.

[0012] According to the patent text structuring and multi-hop semantic query method provided by the present invention, after constructing the patent knowledge graph, the method further includes: Identify and delete isolated nodes and target subgraphs in the patent knowledge graph whose number of entities is below a set threshold to obtain the governed patent knowledge graph.

[0013] According to the patent text structuring and multi-hop semantic query method provided by the present invention, after constructing the patent knowledge graph, the method further includes: A knowledge graph update mechanism is established under the conditions of patent addition and deletion, and the patent knowledge graph is updated according to the knowledge graph update mechanism to obtain the updated patent knowledge graph.

[0014] According to the patented text structuring and multi-hop semantic query method provided by this invention, after receiving the user's query question, the user's query intent type is determined, including: Semantic parsing and intent recognition are performed on the query to obtain probability values ​​of analytical summary intent and statistical query intent. If the probability value of the analysis and summary type intent is above the set probability threshold, then it is determined that the user query intent type includes the analysis and summary type. If the probability value of the statistical query intent is above a set probability threshold, then it is determined that the user's query intent type includes the statistical query type. If the probability value of the analysis and summary intent is lower than the set probability threshold, and the probability value of the statistical query intent is lower than the set probability threshold, then the user query intent type is determined to be other types.

[0015] According to the patent text structuring and multi-hop semantic query method provided by the present invention, based on the user query intent type, the relational database and / or the patent knowledge graph are invoked to perform multi-hop semantic queries to obtain semantic query results, including: If the user's query intent type includes the analysis and summary type, then the analysis and summary question in the query question is extracted, and based on the analysis and summary question, the patent knowledge graph is retrieved for information retrieval to obtain the key information of the conclusion; If the user query intent type includes a statistical query type, then the statistical query question in the query question is extracted, and based on the statistical query question, the relational database is retrieved to obtain the statistical data results; If the user's query intent type is other types, then based on the question to be queried, a pre-established general question and answer database is invoked to retrieve information and obtain the answer to be output.

[0016] On the other hand, the present invention also provides a patented text structuring processing and multi-hop semantic query system, comprising: The acquisition module is used to acquire the patent documents to be processed, perform structured cleaning and format regularization on the patent documents to be processed, generate standardized input text, and store the key structured information obtained from the structured cleaning into a relational database. The construction module is used to construct a patent knowledge graph based on the standardized input text by combining entity and relationship recognition with clustering and global indexing. The query module is used to determine the user's query intent type after receiving the user's query question, and to perform multi-hop semantic query based on the user's query intent type by calling the relational database and / or the patent knowledge graph to obtain the semantic query result; The feedback module is used to generate and return the answer to the query to the user based on the semantic query results.

[0017] The patent text structuring and multi-hop semantic query method and system provided by this invention performs targeted structuring cleaning and format regularization on the patent documents to be processed, generating standardized input text and storing key structured information in a relational database, laying a high-quality and standardized data foundation for subsequent query analysis. Simultaneously, it constructs a patent knowledge graph based on the standardized input text, combining the structured data of the relational database with the semantic association characteristics of the knowledge graph. Based on the user's query intent type, it accurately calls corresponding data resources to conduct multi-hop semantic queries, achieving an organic integration of structured statistical query and deep semantic association reasoning, significantly improving the accuracy, comprehensiveness, and reliability of patent query results. Furthermore, by accurately diverting query tasks through intent recognition and integrating results to generate standardized answers for users, it optimizes the interactive experience and query efficiency of patent retrieval, providing accurate and efficient patent information retrieval support for patent analysis, R&D innovation planning, and technology layout assessment. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating the patented text structuring and multi-hop semantic query method provided in an embodiment of the present invention. Figure 2 This is a flowchart illustrating the process of cleaning and structuring patent texts. Figure 3 This is a flowchart illustrating the process of constructing a patent knowledge graph; Figure 4 This is a schematic diagram illustrating the implementation process of the new patent's map update mechanism; Figure 5 This is a schematic diagram illustrating the implementation process of the map update mechanism corresponding to the deletion of patents; Figure 6 This is a schematic diagram illustrating the implementation process of the intelligent question-and-answer process in patent archives; Figure 7 This is a schematic diagram of the structure of the patented text structuring and multi-hop semantic query system provided in the embodiments of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0021] The following is combined Figures 1 to 7 This invention describes the detailed scheme of the patented text structuring and multi-hop semantic query method and system provided in the embodiments of the present invention.

[0022] like Figure 1 As shown, the patented text structuring and multi-hop semantic query method provided in this embodiment of the invention mainly includes the following steps: Step 110: Obtain the patent document to be processed, perform structured cleaning and format regularization on the patent document to be processed, generate standardized input text, and store the key structured information obtained from the structured cleaning into a relational database.

[0023] Understandably, by performing structured cleaning and format standardization on the patent documents to be processed, the layout noise and redundant information in the original patent documents can be effectively removed, and standardized input text can be generated, laying a high-quality and standardized text foundation for the accurate construction of the subsequent patent knowledge graph. At the same time, the key structured information obtained from the cleaning is stored in a relational database, realizing the structured and standardized management of the core patent information. This not only facilitates accurate statistical queries in the future, but also provides reliable data support for the efficient retrieval, management and maintenance of patent data.

[0024] Step 120: Based on the standardized input text, a patent knowledge graph is constructed by combining entity and relationship recognition with clustering and global indexing.

[0025] This embodiment relies on standardized input text and mines the core technical connections in patent texts through entity and relationship recognition. It combines clustering and global indexing to construct a patent knowledge graph, transforming the unstructured information of patent texts into a structured entity relationship network, thus realizing the visualization and relational expression of patent knowledge. This patent knowledge graph can capture deep technical semantic connections between patents, providing a structured knowledge carrier for subsequent multi-hop semantic queries. It effectively solves the problem that traditional retrieval is difficult to achieve cross-document and cross-entity multi-hop reasoning, and improves the depth and comprehensiveness of patent knowledge mining.

[0026] Step 130: After receiving the user's query question, determine the user's query intent type, and based on the user's query intent type, call the relational database and / or patent knowledge graph to perform multi-hop semantic query to obtain the semantic query result.

[0027] This embodiment achieves precise task allocation and efficient resource utilization for patent search by determining the user's query intent type and selectively calling relational databases and patent knowledge graphs to conduct multi-hop semantic queries. It matches specific query vehicles to different intents, allowing statistical questions to obtain accurate structured data from relational databases and analytical questions to achieve deep multi-hop semantic reasoning using knowledge graphs. This effectively balances the needs for accurate statistics and in-depth analysis in patent searches, significantly reducing the probability of missed and false detections in traditional retrieval methods and improving the accuracy and targeting of multi-hop semantic queries.

[0028] Step 140: Based on the semantic query results, generate and return the answer to the query to the user.

[0029] This embodiment generates and returns standardized answers based on semantic query results, which can integrate and standardize various query results, ensuring the logic, professionalism, and completeness of the returned answers. By directly outputting accurate answers to users' questions, it simplifies the process of users filtering and sorting patent information, optimizes the interactive experience and query efficiency of patent retrieval, and allows users to quickly obtain the patent information they need. This enables the value of patent data to be released efficiently, and provides direct and reliable information references for R&D innovation analysis, technology layout planning, and other work.

[0030] In one embodiment, combined with Figure 2 As shown, the patent documents to be processed undergo structured cleaning and format standardization to generate standardized input text, specifically including: First, the patent documents to be processed are initially cleaned and converted to the desired format to obtain documents in the specified format.

[0031] In practical applications, during the process of inputting patent documents to be processed, patent file managers can upload PDF format patent documents to be processed into the patent file management system. This supports single or batch uploads. The above process corresponds to... Figure 2 Step 210.

[0032] In this embodiment, after reading the patent document to be processed, preliminary cleaning can be performed. Specifically, for the patent document in PDF format, text can be extracted and preliminarily cleaned based on PyMuPDF's layout parsing technology. This involves identifying the authorization announcement number and date, inserting them into the metadata, and then progressively cleaning up layout noise in the patent document, merging forcibly broken Chinese lines, and processing patent metadata (such as patentee, invention title, and abstract). This achieves noise reduction and deletion of layout junk information, ultimately outputting a structurally stable and semantically clear document in a predefined format, specifically a TXT format document, suitable for further parsing or information extraction. The above process corresponds to... Figure 2Step 220 in the process.

[0033] Then, the document with the specified format is structurally segmented to identify several key blocks.

[0034] In this embodiment, the entire text can be segmented according to its structure. Specifically, regular expressions can be used to parse and split the full patent text, automatically identifying and separating the bibliographic data, claims, and specification into three parts, thereby correcting common text concatenation problems. The above process corresponds to... Figure 2 Step 230 in the process.

[0035] Next, deep cleaning is performed on the content of each key block to obtain the cleaned content of each key block.

[0036] In a specific implementation, several key blocks include: bibliographic entries, claims, and specification.

[0037] Furthermore, deep cleaning is performed on the content of each key block to obtain the cleaned content of each key block, specifically including: On the one hand, key fields in the bibliographic project are identified and cleaned, and special rules are applied to the multiple target contents of the key fields to obtain a bibliographic structure dictionary, which is used as the cleaned content of the bibliographic project.

[0038] Specifically, in the deep cleaning stage of bibliographic data, further structured extraction of bibliographic data can be performed. Regular expressions are used as the primary tool to identify and clean key fields from standardized input text, such as application number, authorization announcement number, inventor, agency, and IPC classification number. Furthermore, specific rule-based processing is applied to multiple target contents, including name lists, agency names, and patent types, ultimately generating a structured bibliographic dictionary. The above process corresponds to… Figure 2 Step 240 in the process.

[0039] On the other hand, the effective content of each claim in the claim book is identified and extracted in segments according to the number, and the effective content of each claim is formatted and cleaned, and a claim list is output as the cleaned content of the claim book.

[0040] Specifically, in the deep cleaning stage of the claims document, each claim can be automatically identified from the text, and the valid content of each claim can be extracted in segments according to its number. The valid content of each claim is then formatted and cleaned, such as merging broken lines, deleting extra spaces, and merging split Chinese characters. Finally, a neat and organized list of claims is output. The above process corresponds to… Figure 2 Step 250 in the process.

[0041] On the other hand, key chapters of the instruction manual are extracted, paragraphs are organized, and the invention content and specific implementation methods in the key chapters are assembled to generate an overview of the instruction manual, which serves as the cleaned-up content of the instruction manual.

[0042] Specifically, the deep cleaning stage of the instruction manual mainly involves text structure parsing, cleaning, and generating a technical solution outline. First, key sections such as the technical field, background technology, invention content, and specific implementation methods are automatically identified and their content extracted. Then, paragraph formatting is performed on the key sections, eliminating invisible characters, redundant whitespace, and abnormal line breaks. The invention content and specific implementation methods are combined into a large model input, and a high-density technical solution outline is generated by calling LLM. This outline is then reassembled with the extracted structural blocks to form a standardized instruction manual overview. The above process corresponds to… Figure 2 Step 260 in the process.

[0043] Finally, the cleaned content of all key blocks is integrated and formatted to generate standardized input text.

[0044] In this embodiment, to facilitate subsequent statistical analysis and querying of patent information, key structured information such as the authorization announcement number, application information, patent type, patentee, inventor, IPC classification number, invention title, and abstract can be organized and stored in a relational database. Simultaneously, a unified data table structure enables standardized management of patent data, providing reliable data support for subsequent analysis and business applications. The above process corresponds to... Figure 2 Step 270 in the process.

[0045] At the same time, it is necessary to integrate the cleaned content of all key blocks, generate standardized input text in TXT format, and store it in the map generation folder. The above process corresponds to... Figure 2 Step 280 in the process.

[0046] In one embodiment, combined with Figure 3 As shown, based on standardized input text, a patent knowledge graph is constructed through entity and relationship recognition combined with clustering and global indexing, specifically including: First, the standardized input text is imported and preprocessed to obtain the preprocessed text.

[0047] In practical applications, the underlying architecture for the construction of patent knowledge graphs can adopt the GraphRAG (Graph-enhanced Retrieval-Augmented Generation) model. GraphRAG is an advanced technology that introduces knowledge graphs and graph structure reasoning on the basis of traditional retrieval enhancement generation technology. By performing structured modeling of entities, relationships and attributes in documents, it achieves an upgrade from keyword matching to multi-hop association reasoning. It can effectively improve the accuracy of information retrieval, semantic understanding ability and logical association analysis level in complex knowledge scenarios. In professional fields such as patent retrieval, scientific and technological archive management and intelligence analysis, it can significantly reduce noise interference and enhance the accuracy and interpretability of answers.

[0048] Before text import and basic preprocessing, the GraphRAG model can be fine-tuned and optimized. Specifically, by rewriting the extract_graph (entity relation extraction) and summarize_descriptions (descriptive information summarization) templates, it achieves strongly constrained entity relation extraction and structured technical semantic summarization for patent scenarios, significantly improving the determinism, professionalism, and reusability of graph generation. Combined with customized settings, the model, concurrency, vectorization, and processing flow are systematically optimized, making GraphRAG run more stably, efficiently, and with strict consistency on large-scale patent corpora.

[0049] The core focus of the GraphRAG model fine-tuning work is to construct a rigorous, highly constrained, and fully deterministic entity-relation generation mechanism for patent texts. By revising the extract_graph.txt configuration file, rules related to entity recognition, field mapping, attribute extraction, relation types, and cross-document entity uniqueness are clearly defined. Anchor-based constraints are particularly applied to PATENT entities to ensure that only one patent node is generated per document, strictly prohibiting any inference, completion, or duplicate creation operations. Simultaneously, the sources of entities and relations, as well as the output format, are strictly limited to guarantee the stability, controllability, and professionalism of the knowledge graph structure.

[0050] The core optimization of the `summarize_descriptions.txt` configuration file is to upgrade the patent abstract generation task from traditional text summarization to precise extraction and structured organization of technical semantics. By clearly defining rules, the system is required to extract five categories of technical semantic elements—components, functions, steps, problems, and effects—from relevant paragraphs such as patent abstracts, claims, and background technology. The extraction process is based entirely on the original text, without any subjective inference. Simultaneously, the output format is strictly standardized to ensure that the extracted results can be directly applied to patent clustering, technology route analysis, and cross-patent comparison. This optimization significantly improves the technical content of patent abstracts, making the semantic descriptions generated by the GraphRAG model more structured, reusable, and better suited to specialized computing and analysis needs. The finely tuned GraphRAG model can then be used to further construct a patent knowledge graph.

[0051] In the text import and basic preprocessing stage, standardized input text can be imported and preprocessed, including unifying encoding formats, validating document structure, and identifying fixed field boundaries, to ensure data integrity, standardized format, and clear structure. This process corresponds to... Figure 3 Step 310 aims to establish a consistent and reliable data starting point for subsequent semantic parsing and entity extraction, thereby improving the overall processing quality and stability.

[0052] Then, according to the inherent structure of the patent document, the preprocessed text is semantically segmented and multiple content units are constructed.

[0053] In the semantic segmentation and content unit construction stage, the preprocessed text can be divided into multiple semantically clear and logically independent content units based on the inherent structure of the patent document, such as the technical field, abstract, and claims. The above process corresponds to... Figure 3 Step 320 helps improve the model's ability to understand technical topics, discourse levels, and field meanings, making subsequent entity recognition, attribute extraction, and relationship judgment more accurate and controllable.

[0054] Next, patent technology entities are extracted from multiple content units, and the entity technology relationships between patent technology entities are identified. Based on the patent technology entities and entity technology relationships, an initial graph structure is constructed.

[0055] In the entity recognition and structured technical information extraction stage, based on semantic segmentation, patent technology entities can be extracted according to preset entity types and field mapping rules. This includes information such as patent subject, inventor, patentee, IPC classification, and technical field. Standardization processing is then performed to ensure reliable entity sources, clear definitions, and consistent naming, providing a high-quality set of basic nodes for graph construction. The above process corresponds to… Figure 3 Step 330 in the process.

[0056] Furthermore, in the entity relationship identification and graph structure generation stage, based on document structure and rule constraints, the entity technical relationships between patent technology entities can be identified and an initial graph structure can be constructed, such as inventor-patent, patent-IPC, patent-technical field, etc. All entity technical relationships must have a clear origin, be logically clear, and be recorded in a unified format, thereby forming a systematic patent knowledge association network. The above process corresponds to... Figure 3 Step 340 in the process.

[0057] Finally, the initial graph structure is post-processed through clustering and global indexing to construct the patent knowledge graph.

[0058] In one specific implementation, the initial graph structure is post-processed through clustering and global indexing to obtain a patent knowledge graph, including: First, the initial graph structure is subjected to topic clustering to obtain semantically related local knowledge units and corresponding topic explanation texts.

[0059] In the topic clustering and knowledge unit integration stage, after the initial entity and relationship construction is completed, the initial graph structure can be analyzed using clustering algorithms to identify semantically related local knowledge units, forming community structures such as patent topics, technical modules, or technical directions. Simultaneously, corresponding topic explanation texts are generated. The above process corresponds to... Figure 3 Step 350 in the process can improve the interpretability and overview of the map, making it easier for subsequent retrieval, analysis and knowledge organization.

[0060] Then, based on the initial graph structure, local knowledge units, and topic explanation text, a patent knowledge graph is generated through global indexing.

[0061] In the global index construction and graph result storage stage, all patent technology entities, entity technology relationships, community structures containing local knowledge units, and relevant semantic summaries of the main explanatory text can be uniformly integrated to generate structured data files such as vector indexes, node tables, and edge tables. Subsequently, according to system configuration requirements, these results are stored in a standardized and orderly manner in a designated directory for subsequent graph queries, semantic retrieval, analysis services, and knowledge applications. The resulting patent knowledge graph has a clear structure and rigorous organization, which is conducive to carrying out business operations such as patent knowledge management and enhanced retrieval generation. The above process corresponds to... Figure 3 Step 360 in the process.

[0062] In one embodiment, after constructing the patent knowledge graph, the above-mentioned patent text structuring and multi-hop semantic query method may further include: Identify and delete isolated nodes and target subgraphs in the patent knowledge graph whose number of entities is below a set threshold to obtain the governed patent knowledge graph.

[0063] In the graph governance and structure optimization stage, after the patent knowledge graph is generated, the overall quality and usability can be further improved by governing and sifting the graph structure. This stage mainly includes identifying and removing isolated nodes and deleting target subgraphs with 2-3 entities, thereby improving graph density and query efficiency. The above process corresponds to... Figure 3 In step 370, the patent knowledge graph after treatment has a more compact and reasonable structure, and its relationship coherence and information carrying capacity are significantly enhanced, thus providing a higher quality data foundation for subsequent search enhancement generation, patent analysis and association mining.

[0064] In one embodiment, after constructing the patent knowledge graph, the above-mentioned patent text structuring and multi-hop semantic query method may further include: Establish a knowledge graph update mechanism under the conditions of patent addition and deletion, and update the patent knowledge graph according to the knowledge graph update mechanism to obtain the updated patent knowledge graph.

[0065] In this embodiment, after a new patent document is added, text cleaning and structuring will be automatically completed, and the modification time will be recorded. Subsequently, it will be determined whether the patent knowledge graph needs to be updated through manual triggering or system-periodized checks. If content changes are detected, the graph index will be updated.

[0066] In practical applications, considering the diverse update scenarios, the map update mechanism includes map update mechanisms for newly added patents and map update mechanisms for deleted patents.

[0067] like Figure 4 As shown, the operation process involved in the map update mechanism of the newly added patent is as follows: Step 410: In the patent PDF upload step, users can click the "Add Patent" option to upload the patent file in PDF format. Uploading a single file or multiple files in batches is supported. After uploading, the patent file will be saved in the pending processing directory.

[0068] Step 420: In the automatic text cleaning and structuring process, the PDF content in the patent document to be processed can be automatically extracted, cleaned, parsed and generated into a corresponding TXT file. At the same time, the modification time of the TXT folder is recorded as modified_time, thus providing a basis for subsequent update judgment.

[0069] Step 430: In the step of determining whether the map update is manually triggered, check whether the user actively clicks the update map button. If the user triggers it manually, the map update process will begin immediately.

[0070] Step 440: The main step is to determine if there are any target files whose modified time (modified_time) is later than the current map construction time (build_time). If so, it means that there is new content to be updated.

[0071] Step 450: Execute the graph index update and update the graph build time (build_time) at the same time.

[0072] Step 460: Trigger a scheduled self-check task. Regardless of whether the user manually triggers the update, a scheduled self-check task is triggered to determine if there is a folder modification time (modified_time) that is later than the current map construction time (build_time). If so, it means that there is new content to be updated, and the update action is executed.

[0073] like Figure 5 As shown, the operation process involved in the spectrum update mechanism for deleting patents is as follows: Step 510: Select the patent entries to be deleted and confirm the deletion. In practical applications, users can select the patent records to be deleted in the interactive interface and confirm the deletion. During this process, the consequences of deletion should be displayed to ensure that the operation is correct before proceeding to the subsequent processing steps.

[0074] Step 520: Delete the corresponding data from the relational database, and remove the relevant data records from the relational database based on the unique identifier of the selected patent, such as the authorization announcement number, to ensure that cascading relationships or reference data are updated synchronously and to avoid generating isolated data.

[0075] Step 530: Delete the TXT file corresponding to the patent, and update the folder's modification time (modified_time) accordingly. Specifically, locate and delete the patent's TXT text file, and simultaneously update the file directory's modification time to ensure the file is not indexed or referenced again after cleanup.

[0076] Step 540: Remove the corresponding PDF document of the patent and add it to the removed files folder. This step can be used for subsequent auditing or recovery to ensure that the file does not appear in the main directory. It should be noted that steps 520, 530, and 540 above need to be completed in parallel.

[0077] Step 550: In the step of determining whether the map update is manually triggered, check whether the user actively clicks the update map button. If the user triggers it manually, the map update process will begin immediately.

[0078] Step 560: The main step is to determine if there are any target files whose modified time (modified_time) is later than the current map construction time (build_time). If so, it means that there is new content to be updated.

[0079] Step 570: Execute the graph index update, and update the graph build time (build_time) at the same time.

[0080] Step 580: Trigger a scheduled self-check task. Regardless of whether the user manually triggers the update, a scheduled self-check task is triggered to determine if there is a folder modification time (modified_time) that is later than the current map construction time (build_time). If so, it means that there is new content to be updated, and the update action is executed.

[0081] In the intelligent question-answering workflow construction phase of patent archives, the workflow conducts intent analysis using a Large Language Model (LLM) to identify the probabilities of statistical queries and analytical summaries. After parallel judgment, for questions with analytical summaries, the core content is extracted and the GraphRAG tool is invoked; for questions with statistical queries, parameters are extracted and SQL statements are generated for execution; for questions not involving either type of intent, a regular question-answering service is invoked. Finally, multi-source information is integrated, and the LLM generates standardized answers to provide feedback to the user, achieving accurate and efficient intelligent interaction support.

[0082] Understandably, large language models are artificial intelligence models trained on massive amounts of text data based on deep learning. They possess powerful natural language understanding, generation, semantic representation, and logical reasoning capabilities. They can understand context, complete tasks such as question answering, summarizing, translating, and text generation, and are the core foundational models for current applications such as knowledge retrieval, intelligent question answering, content understanding, and decision support.

[0083] In one embodiment, combined with Figure 6 As shown, after receiving the user's query question, the user's query intent type is determined, specifically including: First, semantic parsing and intent recognition are performed on the query question to obtain the probability values ​​of analytical summary intent and statistical query intent.

[0084] In this embodiment, during the user question phase, the system can receive the user's query questions. In practical applications, users can submit questions related to patent files through the interactive interface, serving as the initial input for the entire process and providing the original basis for subsequent intent analysis and task processing. This process corresponds to... Figure 6 Step 610 in the process.

[0085] In the user intent analysis phase of LLM, a large language model can be used to perform semantic parsing and intent recognition on the user's query, outputting probability values ​​for analytical summary and statistical query-type intents. These probability values ​​provide quantitative basis for parallel branch decisions. This process corresponds to... Figure 6 Step 620 in the process.

[0086] In one scenario, if the probability value of the analysis and summary intent is above a set probability threshold, then it is determined that the user's query intent type includes the analysis and summary type.

[0087] In the intent analysis and summary judgment stage, the probability value of the intent analysis and summary can be compared with a set probability threshold to determine whether the user's question contains an intent analysis and summary. If it is determined to be so, the intent analysis and summary processing flow is initiated. This process corresponds to... Figure 6 Step 630 in the process.

[0088] In another scenario, if the probability value of a statistical query intent is above a set probability threshold, then it is determined that the user's query intent type includes a statistical query type.

[0089] In the statistical query intent determination stage, the probability value of statistical query intent can be compared with a set probability threshold to determine whether the user's question contains statistical query intent. If it is determined to be so, the statistical query intent processing flow is initiated. This process corresponds to... Figure 6 Step 640 in the process.

[0090] In another scenario, if the probability value of the analysis and summary intent is lower than the set probability threshold, and the probability value of the statistical query intent is also lower than the set probability threshold, then the user's query intent type is determined to be other types.

[0091] In other intent determination stages, if, based on the above determinations, the user's question does not contain either an analytical or statistical query intent, it is classified as another type of intent and enters the corresponding processing flow. Figure 6 Step 650 in the middle.

[0092] In one embodiment, based on the user's query intent type, a relational database and / or patent knowledge graph are invoked to perform a multi-hop semantic query to obtain semantic query results, specifically including: In one scenario, if the user's query intent includes an analysis and summary type, the analysis and summary question is extracted from the query question, and based on the analysis and summary question, the patent knowledge graph is retrieved for information retrieval to obtain the key information of the conclusion.

[0093] In this scenario, the questions can be extracted, analyzed, and summarized, while irrelevant content can be filtered out. For user questions containing analytical or summarizing intent, semantic extraction technology can be used to accurately locate and summarize the core questions, filtering out redundant and irrelevant expressions to ensure that subsequent processing focuses on valid information. This process corresponds to... Figure 6 Step 660 in the process.

[0094] Furthermore, statistical analysis tools based on patent knowledge graph-enhanced retrieval (GraphRAG) can be invoked to perform deep retrieval and correlation analysis on the analysis and summary questions extracted in step 660, thereby obtaining key information for the conclusions. This process corresponds to... Figure 6 Step 670 in the process.

[0095] In another scenario, if the user's query intent includes a statistical query type, the statistical query question is extracted from the query question, and the relational database is retrieved based on the statistical query question to obtain the statistical data results.

[0096] In this scenario, the statistical query question can be extracted, and irrelevant content can be filtered out. For user questions containing statistical query intent, semantic extraction technology can be used to accurately locate the core statistical query question, filtering out redundant and irrelevant expressions to ensure that subsequent queries focus on valid needs. This process corresponds to... Figure 6 Step 680 in the process.

[0097] Furthermore, query parameters can be parsed from the extracted statistical query question, a relational database can be called to generate structured SQL statements, the query can be executed, and accurate statistical data results can be returned. This process corresponds to... Figure 6 Step 690 in the process.

[0098] In practical applications, information from analytical summaries and statistical query results can be integrated. The processing results of analytical summaries and statistical query intents can be merged, and LLM (Local Management Model) can be applied to logically organize and generate content from the integrated information. Answers to the questions to be processed are then formed according to a standardized format and fed back to the user. This process corresponds to... Figure 6 Step 6100 in the process.

[0099] In another scenario, if the user's query intent is of another type, the system will retrieve information from a pre-established general question and answer database based on the question being queried, and then output the answer content.

[0100] In this scenario, if the user's question does not contain any analytical, summarizing, or statistical query intent, it is determined to be of another intent type. In other words, the user's query intent type is "other." At this point, a regular question-and-answer service can be invoked, leveraging the LLM and knowledge base to generate the corresponding answer and provide it to the user. This process corresponds to... Figure 6 Step 6110 in the process.

[0101] Based on the same general inventive concept, this invention also protects a patented text structuring and multi-hop semantic query system. The patented text structuring and multi-hop semantic query system provided by this invention will be described below. The patented text structuring and multi-hop semantic query system described below and the patented text structuring and multi-hop semantic query method described above can be referred to in correspondence.

[0102] like Figure 7 As shown, the patented text structuring and multi-hop semantic query system provided in this embodiment of the invention specifically includes: The acquisition module 710 is used to acquire the patent documents to be processed, perform structured cleaning and format regularization on the patent documents to be processed, generate standardized input text, and store the key structured information obtained from the structured cleaning into a relational database.

[0103] Module 720 is used to construct a patent knowledge graph based on standardized input text by combining entity and relationship recognition with clustering and global indexing.

[0104] The query module 730 is used to determine the user's query intent type after receiving the user's query question, and to call the relational database and / or patent knowledge graph to perform multi-hop semantic query based on the user's query intent type, so as to obtain the semantic query result.

[0105] The feedback module 740 is used to generate and return the answer to the query to the user based on the semantic query results.

[0106] Regarding the system in the above embodiments, the specific ways in which each module performs operations have been described in detail in the embodiments of the relevant methods, and will not be elaborated further here.

[0107] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A patented text structuring processing and multi-hop semantic query method, characterized in that, include: Obtain the patent document to be processed, perform structured cleaning and format regularization on the patent document to be processed, generate standardized input text, and store the key structured information obtained from the structured cleaning into a relational database; Based on the standardized input text, a patent knowledge graph is constructed by combining entity and relationship recognition with clustering and global indexing. After receiving the user's query question, the query intent type is determined, and based on the query intent type, the relational database and / or the patent knowledge graph are invoked to perform a multi-hop semantic query to obtain the semantic query result. Based on the semantic query results, the answer to the query question is generated and returned to the user.

2. The patent text structuring processing and multi-hop semantic query method according to claim 1, characterized in that, The patent documents to be processed are subjected to structured cleaning and format standardization to generate standardized input text, including: The patent documents to be processed are initially cleaned and converted to a specific format to obtain a document in the set format. The document with the specified format is structurally segmented to identify multiple key blocks; Deep cleaning is performed on the content of each key block to obtain the cleaned content of each key block; The cleaned content of all key blocks is integrated and formatted to generate standardized input text.

3. The patent text structuring processing and multi-hop semantic query method according to claim 2, characterized in that, The key sections include: bibliographic information, claims, and specification. Deep cleaning is performed on the content of each key block to obtain the cleaned content of each key block, including: Identify and clean the key fields in the bibliographic data, perform special rule-based processing on the multiple target contents of the key fields, and obtain a bibliographic structure dictionary, which is used as the cleaned content of the bibliographic data. Identify and extract the valid content of each claim in the claim book by segment according to the number, and perform format standardization and text cleaning on the valid content of each claim, output the claim list, and use it as the cleaned content of the claim book; Key sections of the specification are extracted, paragraphs are organized, and the invention content and specific embodiments in the organized key sections are assembled to generate an overview of the specification, which serves as the cleaned-up content of the specification.

4. The patent text structuring processing and multi-hop semantic query method according to claim 1, characterized in that, Based on the standardized input text, a patent knowledge graph is constructed by combining entity and relation recognition with clustering and global indexing, including: The standardized input text is subjected to text import and basic preprocessing to obtain preprocessed text; According to the inherent structure of the patent document, the preprocessed text is semantically segmented and multiple content units are constructed. Extract the patent technology entities from the multiple content units, identify the entity technology relationships between the patent technology entities, and construct an initial graph structure based on the patent technology entities and the entity technology relationships; By clustering and global indexing, the initial graph structure is post-processed to construct a patent knowledge graph.

5. The patent text structuring processing and multi-hop semantic query method according to claim 4, characterized in that, By performing clustering and global indexing on the initial graph structure, a patent knowledge graph is obtained, including: The initial graph structure is subjected to topic clustering to obtain semantically related local knowledge units and corresponding topic explanation texts; Based on the initial graph structure, the local knowledge units, and the topic explanation text, a patent knowledge graph is generated through global indexing.

6. The patent text structuring processing and multi-hop semantic query method according to any one of claims 1 to 5, characterized in that, After constructing the patent knowledge graph, the method further includes: Identify and delete isolated nodes and target subgraphs in the patent knowledge graph whose number of entities is below a set threshold to obtain the governed patent knowledge graph.

7. The patent text structuring processing and multi-hop semantic query method according to any one of claims 1 to 5, characterized in that, After constructing the patent knowledge graph, the method further includes: A knowledge graph update mechanism is established under the conditions of patent addition and deletion, and the patent knowledge graph is updated according to the knowledge graph update mechanism to obtain the updated patent knowledge graph.

8. The patent text structuring processing and multi-hop semantic query method according to claim 1, characterized in that, After receiving the user's query question, determine the type of user's query intent, including: Semantic parsing and intent recognition are performed on the query to obtain probability values ​​of analytical summary intent and statistical query intent. If the probability value of the analysis and summary type intent is above the set probability threshold, then it is determined that the user query intent type includes the analysis and summary type. If the probability value of the statistical query intent is above a set probability threshold, then it is determined that the user's query intent type includes the statistical query type. If the probability value of the analysis and summary intent is lower than the set probability threshold, and the probability value of the statistical query intent is lower than the set probability threshold, then the user query intent type is determined to be other types.

9. The patent text structuring processing and multi-hop semantic query method according to claim 8, characterized in that, Based on the user's query intent type, a multi-hop semantic query is performed by invoking the relational database and / or the patent knowledge graph to obtain semantic query results, including: If the user's query intent type includes the analysis and summary type, then the analysis and summary question in the query question is extracted, and based on the analysis and summary question, the patent knowledge graph is retrieved for information retrieval to obtain the key information of the conclusion; If the user query intent type includes a statistical query type, then the statistical query question in the query question is extracted, and based on the statistical query question, the relational database is retrieved to obtain the statistical data results; If the user's query intent type is other types, then based on the question to be queried, a pre-established general question and answer database is invoked to retrieve information and obtain the answer to be output.

10. A patented text structuring processing and multi-hop semantic query system, characterized in that, include: The acquisition module is used to acquire the patent documents to be processed, perform structured cleaning and format regularization on the patent documents to be processed, generate standardized input text, and store the key structured information obtained from the structured cleaning into a relational database. The construction module is used to construct a patent knowledge graph based on the standardized input text by combining entity and relationship recognition with clustering and global indexing. The query module is used to determine the user's query intent type after receiving the user's query question, and to perform multi-hop semantic query based on the user's query intent type by calling the relational database and / or the patent knowledge graph to obtain the semantic query result; The feedback module is used to generate and return the answer to the query to the user based on the semantic query results.

Images