A patent question and answer method and system based on knowledge graph enhanced retrieval

Abstract

The application discloses a patent question and answer method and system based on knowledge graph enhanced retrieval, and belongs to the technical field of artificial intelligence and patent information processing. The method carries out standardization processing and structured technical semantic extraction on patent data, and constructs an initial patent knowledge graph; based on vector similarity and patent classification number compatible constraint normalization technical entity, a three-layer patent knowledge graph and a mixed retrieval index containing sparse index, dense vector index and relationship path index are formed; according to a user question, an execution plan is generated and routed to a single patent question and answer or a cross patent analysis track, after evidence acquisition, graph expansion, verification and at most one controlled remedy, a question and answer result with an evidence path is generated. The application can reduce entity redundancy, improve cross patent statistical stability and result reviewability.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence, natural language processing, knowledge graphs, information retrieval, and patent information processing, specifically to a patent question-answering method, system, electronic device, and computer-readable storage medium based on knowledge graph-enhanced retrieval. Background Technology

[0002] With the continuous growth of technological innovation activities, patent documents have become an important information resource for recording technological development paths, reflecting industrial competitive landscapes, and supporting technical intelligence analysis. Patent documents are typically technical, legal, and normative in nature, with lengthy texts, dense terminology, and rigorous sentence structures. Technical information is scattered across multiple parts, including the title, abstract, claims, specification, and figures. These characteristics create a high technical barrier to the automated understanding, retrieval, analysis, and question-answering applications of patent documents.

[0003] Existing patent search systems primarily rely on keyword searches, classification number searches, or field combination searches. While these methods are effective when users have specific keywords, patent numbers, or classification numbers, they are less effective when users ask questions in natural language, such as what technical problem a patent solves, what the main solutions are for a particular technology, or what evolutionary trends a certain technology exhibits over time. Traditional search systems typically only return a list of documents, making it difficult to directly organize and form an interpretable answer.

[0004] Search-enhanced generation techniques combine external document retrieval results with language model generation capabilities for knowledge-intensive question-answering tasks. However, directly applying ordinary search-enhanced generation frameworks to the patent field still has shortcomings: on the one hand, patent texts contain many synonyms, near-synonyms, abbreviations, and domain terms, making it easy to miss, falsely recall, or semantic drift if relying solely on keyword or vector retrieval; on the other hand, the goals of patent question answering are not fixed, ranging from precise explanations of a single patent to statistical analysis, topic aggregation, technological evolution, or trend analysis of multiple patents. A unified search path cannot simultaneously meet the needs of both local factual question answering and global analytical question answering.

[0005] Knowledge graphs can represent objects such as patents, applicants, rights holders, inventors, classification numbers, years, technical problems, solutions, technical effects, core technical themes, and industry sectors as nodes and relationships, thus providing a foundation for structured queries, path reasoning, and cross-document aggregation. However, if a knowledge graph only records patent metadata without extracting technical semantics, it will be difficult to answer questions about the technical substance. If the extracted results are directly added to the graph without entity normalization, a large number of fragmented nodes will easily be generated. For example, semantically similar entities such as "image recognition," "picture recognition," and "video recognition" will be stored in a scattered manner, which will affect the statistical and analytical results.

[0006] Therefore, there is a need for a patent question-answering method that can comprehensively utilize patent knowledge graphs, hybrid search indexes, entity normalization, and task adaptive routing, so that the system can provide accurate, complete, and interpretable technical analysis for a single patent, as well as provide stable statistical, trend analysis, and topic summarization results for cross-patent issues. Summary of the Invention

[0007] The purpose of this invention is to provide a patent question-answering method and system based on knowledge graph-enhanced retrieval, which addresses at least the following problems existing in the prior art: the mixed granularity of patent question-answering tasks leads to incompatible processing paths; patent technical knowledge is implicitly expressed in natural language and lacks structured organization; single keyword retrieval or single vector retrieval is difficult to achieve both precise matching and semantic recall; the fragmentation of synonymous or near-synonymous technical entities affects cross-patent statistics; the lack of graph path evidence in search results leads to insufficient interpretability and verifiability of answers; and the lack of a controlled remedial mechanism when search results are empty, graph paths are missing, or statistics are abnormal.

[0008] Technical solution To achieve the above objectives, the present invention provides a patented question-answering method based on knowledge graph-enhanced retrieval, the method being executed by an electronic device and comprising the following steps.

[0009] S1. Obtain the patent data to be processed, perform basic information processing and structured technical semantic extraction on the patent data to be processed, obtain entity relationship data containing basic patent information, technical entities and technical relationships, and construct an initial patent knowledge graph based on this.

[0010] S2. Based on vector similarity and patent classification number compatibility constraints, the technical entities under the same entity type are normalized to establish a mapping relationship from the original technical entities to the normalized technical entities. Based on the initial patent knowledge graph and the mapping relationship, a three-layer patent knowledge graph including a patent basic information layer, a technical semantic entity layer and a normalized entity layer is constructed, as well as a hybrid retrieval index including sparse index, dense vector index and relation path index.

[0011] S3. Receive a user's natural language question, perform intent recognition on the question to generate a structured execution plan, and route the user's natural language question to a single patent question-and-answer track or a cross-patent analysis track accordingly.

[0012] S4. In the routed track, candidate evidence is obtained based on the hybrid retrieval index and a one-hop or two-hop graph expansion is performed to obtain an evidence set; the evidence set is verified for completeness, consistency and relevance, and at most one controlled remedy retrieval is performed to update the evidence set when a remedyable problem is detected.

[0013] S5. Generate patent question-and-answer results based on the verified evidence set, wherein the patent question-and-answer results include the answer content and the evidence path supporting the answer content.

[0014] Furthermore, the patent data to be processed includes publication number, title, abstract, and patent metadata. The patent metadata includes at least one of the following: applicant, current patentee, inventor, IPC classification number, cited patents, application year, publication date, and grant status.

[0015] Furthermore, the structured semantic extraction is performed through a structured semantic extraction model, and the extraction results include a public number field, an entity array, and a relation array; each entity in the entity array includes an entity name, an entity type, and an entity description, and each relation in the relation array includes a source node, a target node, a relation type, and a relation description.

[0016] Furthermore, the entity type belongs to a preset entity type set, which includes at least one of technical problem entities, solution entities, technical effect entities, core technology theme entities, and industry domain entities; the relationship type belongs to a preset relationship type set, which includes at least one of proposal relationship, adoption relationship, achievement relationship, involvement relationship, resolution relationship, realization relationship, and domain-related patent relationship.

[0017] Furthermore, the relational path index is used to perform path queries, neighborhood expansion, and evidence path organization based on nodes and edges in the graph database, so that the retrieval results not only include text fragments or candidate patents, but also verifiable combinations of nodes, relations, and paths.

[0018] Beneficial effects (1) This invention routes questions through single-patent question-and-answer tracks and cross-patent analysis tracks, so that patent question-and-answer tasks of different granularities adopt matching retrieval, expansion and answer organization strategies, avoiding performance trade-offs caused by using the same retrieval strategy for different tasks.

[0019] (2) This invention transforms the core technical knowledge in the patent text into a searchable, statistical, and reasonable structured representation by pre-defined entity types such as technical problem entity, solution entity, technical effect entity, core technical theme entity and industry domain entity.

[0020] (3) This invention reduces the redundancy of knowledge graph nodes caused by synonyms, near-synonyms and entities with inconsistent granularity by combining the entity normalization mechanism of vector similarity and patent classification number compatibility constraint, and improves the stability of cross-patent statistics and topic aggregation.

[0021] (4) This invention uses a hybrid retrieval system of sparse index, dense vector index and relational path index, combined with sorting fusion and graph expansion, to take into account keyword accuracy, semantic similarity recall and graph structure evidence organization, thereby improving recall coverage and answer reliability in complex patent questions and answers.

[0022] (5) This invention enables answers to carry nodes, relationships, paths, matching methods and supporting text through relational path indexing and structured evidence items, making it easier for users to review and reducing the risk of unverified text directly participating in answer generation.

[0023] (6) This invention uses evidence integrity, consistency and relevance verification and at most one controlled remedy retrieval mechanism to make limited corrections when there are empty technical chains, missing graph paths, abnormal zero value statistics, off-topic candidate evidence or insufficient answer evidence, so as to avoid semantic drift caused by unlimited expansion and reduce retrieval computation overhead. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the five-stage overall process of the patent question-answering method based on knowledge graph-enhanced retrieval in an embodiment of the present invention.

[0025] Figure 2 This is a schematic diagram of a three-layer patent knowledge graph structure in an embodiment of the present invention.

[0026] Figure 3 This is a schematic diagram of a single-patent question-and-answer process in an embodiment of the present invention.

[0027] Figure 4 This is a schematic diagram of the cross-patent analysis track process in an embodiment of the present invention.

[0028] Figure 5 This is a schematic diagram of the evidence verification and controlled remedy process in an embodiment of the present invention.

[0029] Figure 6 This is a schematic diagram of the module composition of the patent question-and-answer system in an embodiment of the present invention.

[0030] Figure 7 This is a schematic diagram of the electronic device structure in an embodiment of the present invention. Detailed Implementation

[0031] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the described embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention. Where there is no conflict, the technical features in the following embodiments can be combined with each other.

[0032] Example 1: Patent Question Answering Method Based on Knowledge Graph Enhanced Retrieval

[0033] This embodiment provides a patented question-answering method based on knowledge graph-enhanced retrieval. This method can be executed by electronic devices, servers, cloud platforms, or distributed computing nodes. The electronic device may include a processor, memory, communication interface, and input / output interface. The knowledge graph can be implemented using a graph database; dense vector indexes can be implemented using a vector database or vector index library; sparse indexes can be implemented using keyword inverted indexes; and relational path indexes can be implemented using path queries, edge type indexes, or pre-computed path indexes from the graph database.

[0034] The method includes the following steps.

[0035] S101. Patent Data Acquisition and Basic Information Processing

[0036] The system first acquires the patent data to be processed. This patent data can come from files exported from a patent database, a structured database, a patent search platform interface, or an enterprise's internal patent management system. Each piece of patent data includes a publication number, title, abstract, and patent metadata. The patent metadata may include at least one of the following: applicant, current patentee, inventor, IPC classification number, cited patents, application year, publication date, and grant status.

[0037] Before extracting semantics using structured technology, the system cleans, splits, deduplicates, and standardizes the publication number, title, abstract, and patent metadata in the patent data to be processed, obtaining basic patent information. Specifically, the system can uniformly format the publication number, removing redundant spaces, illegal characters, and inconsistencies between full-width and half-width characters; separate and deduplicate multi-value fields such as applicant, current rights holder, inventor, and cited patents; perform regular expression matching and hierarchical truncation on the IPC classification number to retain the complete classification number and its major category, minor category, or main group information; and perform numerical or date format validation on time fields such as application year and publication date.

[0038] In one embodiment, the system sets a unique constraint on the publication number of the Patent node, sets name or encoding constraints on the Organization node, Person node, IPC node, and Year node respectively, and creates nodes and relationships using a merged write method. Therefore, even if patent data is imported in multiple batches or repeatedly imported, no duplicate basic information nodes will be generated.

[0039] S102, Structured semantic extraction

[0040] The system performs structured technical semantic extraction based on the patent text information in the patent data to be processed. The patent text information may include the title, abstract, claims, specification text, or at least one of these. Structured technical semantic extraction can be performed by a structured semantic extraction model, a large language model, a rule-based model, or a combination of these models and rules.

[0041] The extraction results include a publication number field, an entity array, and a relation array. Each entity in the entity array includes an entity name, entity type, and entity description; each relation in the relation array includes a source node, a target node, a relation type, and a relation description. The entity type is limited to a preset entity type set, which includes at least one of the following: technical problem entities, solution entities, technical effect entities, core technology theme entities, and industry domain entities.

[0042] The technical problem entity is used to represent the defects of existing technology, difficulties to be solved, or constraints; the solution entity is used to represent the specific technical means, methods, devices, systems, or processes proposed in the patent; the technical effect entity is used to represent the beneficial effects produced after adopting the solution; the core technology theme entity is used to represent the underlying technical terms, key algorithms, key mechanisms, or key components; and the industry field entity is used to represent the application scenario, industry field, or industry direction to which the patent belongs.

[0043] The preset set of relationship types includes: the relationship of patent node pointing to the technical problem entity for proposing a problem; the relationship of patent node pointing to the solution entity for adopting a solution; the relationship of patent node pointing to the technical effect entity for achieving a technical effect; the relationship of patent node pointing to the core technology theme entity for involving a technical problem entity; the relationship of technical problem entity pointing to the solution entity for solving a problem; the relationship of solution entity pointing to the technical effect entity for realizing a technical effect; and the relationship of industry domain entity pointing to patent node related patents.

[0044] In one implementation, the structured semantic extraction model can output the following structured results: { "publication_no": "CNXXXX", "entities": [ {"name": "Technical Problem A", "type": "Technical Problem Entity", "description": "Represents the technical defect to be resolved"}, {"name": "Solution B", "type": "Solution Entity", "description": "Indicates the technical means used to address technical problem A"}, {"name": "Technical Effect C", "type": "Technical Effect Entity", "description": "Represents the technical effect produced by Solution B"} ], "relations": [ {"source": "Technical Problem A", "target": "Solution B", "type": "Relationship to Solution", "description": "Solution B is used to solve Technical Problem A"}, {"source": "Solution B", "target": "Technical Effect C", "type": "Relationship Implementation", "description": "Solution B Achieves Technical Effect C"} ] } S103. Extraction Result Validation and Fault Tolerance

[0045] Because the output of the structured semantic extraction model may have problems such as missing fields, incorrect field types, out-of-bounds enumeration, non-existent relation endpoints, or format that does not conform to the preset structure, the system performs format repair, type verification, enumeration verification, and graph consistency verification after extraction.

[0046] Formatting correction is used to complete fields, delete illegal fields, convert field types, or repair the structure of the extracted results. Type validation is used to confirm that fields such as entity name, entity type, entity description, source node, target node, relation type, and relation description conform to the preset data model. Enumeration validation is used to confirm that both entity type and relation type belong to the preset set.

[0047] Graph consistency verification includes enumerating consistency checks on entity types and relationship types; verifying the existence of source and target nodes for each technical relationship; marking or deleting isolated entities that do not form any relationship with patent nodes or other technical entities; and, in the event of verification failure, performing a re-extraction or field-level repair on the extraction results. The field-level repair does not change the preset entity type set and the preset relationship type set.

[0048] Through the above verification, the system can reduce the impact of illegal nodes, illegal relationships, and unparseable text on subsequent graph construction, hybrid retrieval, and question-answer generation before the extracted results are stored in the database.

[0049] S104, Initial Patent Knowledge Graph Construction

[0050] The system constructs an initial patent knowledge graph based on basic patent information and entity relationship data. This initial patent knowledge graph includes a basic patent information layer and a technical semantic entity layer.

[0051] The basic patent information layer includes Patent nodes, Organization nodes, Person nodes, IPC nodes, and Year nodes. Patent nodes are used to store attributes such as publication number, title, abstract, application year, publication date, and grant status; Organization nodes are used to represent the applicant or current rights holder; Person nodes are used to represent the inventor; IPC nodes are used to represent the patent classification code; and Year nodes are used to represent the application year or publication year.

[0052] In the basic patent information layer, the Patent node and the Organization node can establish an applicant relationship or a current rights holder relationship; the Patent node and the Person node can establish an inventor relationship; the Patent node and the IPC node can establish a classification number relationship; the Patent node and the cited patent Patent node can establish a citation relationship; and the Patent node and the Year node can establish a year relationship.

[0053] The technical semantic entity layer includes technical entity nodes, as well as the HAS_TECH_ENTITY relationship between patent nodes and technical entity nodes, and the TECH_RELATION relationship between different technical entity nodes. The attributes of a technical entity node include at least entity name, entity type, and entity description. The HAS_TECH_ENTITY relationship is used to characterize the association between a patent node and its related technical entities, while the TECH_RELATION relationship is used to characterize the semantic dependency or technical contribution relationship between different technical entities.

[0054] S105, Technical Entity Standardization and Standardized Entity Layer

[0055] The entity normalization module addresses the issue of the same technical concept appearing under different names in different patents. The system first groups technical entities according to entity type, then performs text embedding encoding on the names of technical entities within the same entity type, resulting in fixed-dimensional vectors. Next, the system retrieves candidate synonyms or near-synonyms based on vector similarity.

[0056] For candidate synonyms or near-synonyms, the system determines whether their respective associated patent sets have the same or compatible IPC classification numbers. Only when the entity types are the same, the vector similarity meets preset conditions, and the IPC classification numbers meet compatibility constraints, will the system merge candidate synonyms or near-synonyms into the same specification technical entity.

[0057] In one implementation, the IPC classification number satisfies the compatibility constraint condition by including: the patents associated with the candidate entities respectively have at least one of the following: the same IPC major class, the same IPC subclass, or the same IPC main group. To avoid the erroneous merging of semantically similar words in different technical fields, the system can use the IPC classification number compatibility constraint as a necessary merging condition in addition to vector similarity. Table 1 shows an example of the technical entity normalization merging determination.

[0058] As shown in Table 1, both "image recognition" and "picture recognition" belong to core technology topics and are associated with the G06V classification number. When the semantic similarity meets a preset threshold, the system merges them into the canonical entity "image recognition". Although "image diagnosis" has some semantic similarity to "image recognition", its associated IPC classification number is A61B, which does not meet the compatibility constraint with G06V. Therefore, the system does not merge it with "image recognition" but retains it as an independent canonical entity "image diagnosis". This reduces the risk of erroneous merging of entities across technical fields.

[0059] Table 1. Examples of Technical Entity Standardization Merger Determination Image recognition Core Technology Theme G06V 0.91 yes Image recognition Image recognition Core Technology Theme G06V 0.89 yes Image recognition Imaging diagnosis Core Technology Theme A61B 0.84 no Imaging diagnosis In one embodiment, the system uses a disjoint-set data structure to maintain equivalence classes of candidate synonyms or near-synonyms, and selects the entity name with the shortest name, highest frequency of occurrence, or highest confidence from each equivalence class as the canonical technical entity name. Subsequently, the system creates a CanonicalTechEntity node and establishes a MAPS_TO relationship between the TechEntity node and the CanonicalTechEntity node.

[0060] The normalized entity layer includes CanonicalTechEntity nodes and MAPS_TO relationships between TechEntity nodes and CanonicalTechEntity nodes. Through the normalized entity layer, the system retains the original extracted expressions and can perform unified calculations based on normalized technical entities in cross-patent statistics, topic summarization, trend analysis, and technology profiling analysis, thereby improving statistical stability and the quality of topic aggregation.

[0061] Ultimately, the system constructs a three-layer patent knowledge graph based on the initial patent knowledge graph and the standardized entity layer. This three-layer patent knowledge graph includes a patent basic information layer, a technical semantic entity layer, and a standardized entity layer, such as... Figure 2 As shown.

[0062] S106. Construction of Hybrid Search Index

[0063] The system constructs a hybrid retrieval index associated with a three-layer patent knowledge graph. This hybrid retrieval index includes a keyword-based sparse index, a text-embedded dense vector index, and a graph database-based relational path index.

[0064] Sparse indexes are used for precise or approximate keyword matching based on keywords, patent titles, abstracts, entity names, canonical entity names, and metadata fields. Dense vector indexes encode user questions, technical entity names, canonical entity names, patent titles, or abstracts into vectors and perform semantic recall based on vector similarity. Relational path indexes are used for path queries, neighborhood expansion, and evidence path organization based on nodes and edges in a graph database.

[0065] Ranking fusion is used to merge and rank sparse and dense search results. In one embodiment, the system employs a reciprocal ranking fusion method, whereby the fusion score for a candidate object is calculated based on its ranking in different search engines and a preset smoothing parameter.

[0066] For a candidate object d, its fusion score Score(d) is calculated based on the ranking of the candidate object in different search result lists and a preset smoothing parameter, satisfying the following: , Where rankᵢ(d) is the ranking of candidate object d in the i-th search result list, and k is a preset smoothing parameter.

[0067] The knowledge graph expansion begins with a Patent node, TechEntity node, or CanonicalTechEntity node, performing one- or two-hop expansion within the three-layer patent knowledge graph to obtain relevant patents, technical entities, standardized technical entities, IPC classifications, applicants, years, and relationship paths. The system can dynamically control the expansion depth based on a structured execution plan to avoid insufficient evidence or over-diffusion.

[0068] S107, User Problem Planning and Track Routing

[0069] The question planning and routing module receives user natural language questions, performs intent recognition on the user's natural language questions, and generates a structured execution plan. The structured execution plan may include at least one of the following: question type, core retrieval entity, graph expansion requirements, retrieval hop count, query variants, execution identifier, and answer organization method.

[0070] The system identifies whether the user's natural language question contains publication numbers, patent titles, statistical terms, trend terms, or analytical terms. When a publication number or a uniquely identifiable patent title is present, the question type is determined to be a single-patent Q&A question, and the system routes it to the single-patent Q&A track. When expressions related to quantity, percentage, statistics, or distribution are present, the question type is determined to be a cross-patent statistical Q&A question, and a corresponding execution identifier for the cross-patent analysis track is generated. When expressions related to trends, evolution, hot topics, development characteristics, or technical routes are present, the question type is determined to be a cross-patent analysis Q&A question, and a corresponding execution identifier for the cross-patent analysis track is generated.

[0071] The system also determines core search entities and graph expansion needs based on the identified technical entity words. Through the above-mentioned trajectory routing, the system can use local graph expansion for precise question answering of single patents, and use topic-level semantic subgraph construction for cross-patent statistics or trend analysis, thereby reducing unnecessary search scope and graph expansion range.

[0072] S108, Single Patent Q&A Track

[0073] like Figure 3 As shown, after routing to the single-patent Q&A track according to the structured execution plan, the system prioritizes precise matching of the target patent based on the publication number or patent title. If the question contains a standard publication number, the system performs precise matching on the publication number attribute of the Patent node; if the question contains a patent title, the system performs precise matching or approximate matching on the title attribute of the Patent node.

[0074] When an exact match fails, the system recalls candidate patents based on a hybrid retrieval index and determines the target patent through sorting and fusion. Once the target patent is determined, the system performs a one-hop or two-hop local graph expansion starting from the Patent node corresponding to the target patent to obtain the technical problems, solutions, technical effects, core technical themes, industry sectors, and their relationship paths associated with the target patent, and adds them to the evidence set.

[0075] When users ask questions such as "What technical problem does this patent solve?", "What solution is adopted?", and "What technical effect is achieved?", the system organizes the answers according to the technical problem, solution, technical effect, core technical theme, and industry field, and displays the evidence path between patent nodes, technical entity nodes, and technical relationships. When users ask questions about the applicant, inventor, classification number, application year, or citation information, the system prioritizes answering based on evidence from the basic patent information layer.

[0076] S109, Cross-Patent Analysis Track

[0077] like Figure 4As shown, after routing to the cross-patent analysis track according to the structured execution plan, the system generates multiple query variants based on the user's natural language question, and obtains search results based on the query variants using a hybrid search index. Subsequently, the system determines the seed patent, seed technology entity, or seed specification technology entity through sorting and fusion.

[0078] The system expands the graph based on seed patents, seed technology entities, or seed specification technology entities, constructing a topic-level semantic subgraph and incorporating an evidence set. The topic-level semantic subgraph can include topic-related Patent nodes, TechEntity nodes, CanonicalTechEntity nodes, IPC nodes, Organization nodes, Year nodes, and their relationship paths.

[0079] For cross-patent statistics, the system first searches for standardized technical entities matching the subject terms at the standardized entity layer, and then backtracks to the original TechEntity node and Patent node through the MAPS_TO relationship to obtain the set of patents related to the subject. If no match is found at the standardized entity layer or the statistical results are abnormal, the system can fall back to the original TechEntity name layer for direct matching, or fall back to the patent title and abstract text layer for matching.

[0080] For the problems of theme summarization and technology profiling analysis, the system statistically analyzes the frequency, number of associated patents, and time distribution of different entity types on the theme-related patent set, and summarizes high-frequency technical issues, main solutions, common technical effects, core technical themes, and industry sectors. The system can also statistically analyze the high-frequency problem-solution-effect chain based on the TECH_RELATION relationship, thereby forming a theme-level technology profile.

[0081] For trend analysis, the system comprehensively considers the frequency of occurrence of standardized technical entities, the number of associated patents, the distribution of application years, the distribution of IPC classifications, and the distribution of applicants to generate technical hotspots, development paths, and stage-by-stage evolution characteristics. Outputs across patent analysis tracks include statistical definitions, hit counts, thematic distribution, trend descriptions, representative patents, evidence paths, and uncertainty explanations.

[0082] S110, Evidence Verification and at Most One Controlled Remedy

[0083] like Figure 5 As shown, before generating an answer, the system verifies the completeness, consistency, and relevance of the evidence set. The remediable issues include at least one of the following: empty technical chains, missing graph paths, outlier zero-value statistics, off-topic candidate evidence, or insufficient evidence for the answer.

[0084] For single-patent Q&A, the system verifies whether the target patent has been successfully located, whether the target patent has related technical entities, whether the required technical problem, solution, or technical effect is missing, and whether the local graph path can support the answer. For cross-patent analysis, the system verifies whether the statistical results are abnormally zero or abnormally large, whether there is sufficient evidence in the topic-level semantic subgraph, whether the candidate evidence deviates from the core search entity, and whether the evidence set covers representative patents.

[0085] When the number of hits of the standard technical entity in cross-patent statistical question and answer is zero, but there are candidate patents in the sparse index or dense vector index, it is determined to be an abnormal zero value statistical error; when the vector similarity between the candidate evidence and the core search entity is lower than the second threshold, and there is no graph path connected to the core search entity, it is determined to be off-topic candidate evidence; when the evidence set lacks any type of evidence in the categories of technical problem, solution, or technical effect, it is determined to be insufficient answer evidence.

[0086] When a remediable issue is detected, the system performs at most one controlled remedial retrieval to update the evidence set. The controlled remedial retrieval includes at least one of the following: narrowing the keyword set, lowering the entity matching threshold, switching candidate seed entities, falling back to the original technical entity matching, falling back to the title or abstract text level matching, and discontinuing the use of expanded terms that are prone to semantic drift.

[0087] Controlled remedies follow the principle of limited retries. Each round of question answering triggers at most one remedy. After a remedy, regardless of its success, the system proceeds to the answer generation stage, outputting search criteria, evidence coverage, or uncertainty statements as necessary. This design improves system robustness, avoids evidence drift caused by repeated expansion, and reduces computational resource consumption caused by unbounded searches.

[0088] S111, Answer Generation

[0089] The answer generation module generates patent question-and-answer results based on the verified evidence set. The patent question-and-answer results include the answer content and the evidence paths supporting that answer content.

[0090] When generating patent question-and-answer results, the system uses structured evidence items as input. Each structured evidence item includes at least one of the following: evidence source type, node identifier, node label, relationship type, relationship path, matching score, matching method, and supporting text. The answer generation module organizes natural language answers based on the structured evidence items to reduce the risk of unverified content being generated directly.

[0091] In precise question-and-answer analysis for a single patent, the results include target patent information, technical problem, solution, technical effect, core technology theme, industry sector, and local graph path. In cross-patent question-and-answer analysis, the results include statistical caliber, number of hits, theme distribution, trend description, representative patents, evidence path, and uncertainty explanation.

[0092] In one embodiment, the evidence path in the answer can be represented as a combination of nodes and relationships, such as "Patent node — Proposal relationship — Technical problem entity — Resolution relationship — Solution entity — Implementation relationship — Technical effect entity". By outputting the evidence path, users can verify the graph structure and candidate evidence sources on which the answer is based.

[0093] As shown in Table 2, the evidence source type indicates whether the evidence originates from the patent basic information layer, the technical semantic entity layer, or the normalized entity layer; the node identifier uniquely identifies the evidence node; the node label indicates the node category; the relationship path indicates the combination of nodes and edges in the evidence path; the matching method indicates the recall method for candidate evidence; and the supporting text records the patent title, abstract, or extracted description supporting the answer. The answer generation module organizes the natural language answer based on the above fields and outputs the corresponding evidence path, enabling users to verify the graph structure and evidence source upon which the answer is based.

[0094] Table 2. Description of Fields in Structured Evidence Types of evidence sources This indicates that the evidence originates from the basic information layer, the technical semantic entity layer, or the normalized entity layer. Technical Semantic Entity Layer Node identifier A unique identifier representing an evidence node Patent:CNXXXX Node Labels Indicates node category Patent、TechEntity、CanonicalTechEntity Relationship Path Represents the combination of nodes and edges in the evidence path. Patent—Relationship of Proposal—Entity of Technical Issue Matching method Indicates the recall method of candidate evidence Sparse search, dense search, relational path index Support text The patent title, abstract, or extracted description that supports the answer. Description of technical effects in the abstract Example 2: Patent Question Answering System Based on Knowledge Graph Enhanced Retrieval

[0095] This invention also provides a patent question-answering system based on knowledge graph-enhanced retrieval, such as... Figure 6 As shown. The system can be deployed on a server, cloud platform, electronic device, or distributed computing node, and is used to perform the steps in the above method embodiments.

[0096] The system includes a graph construction module, an entity normalization and hybrid indexing module, an intent routing module, a retrieval verification and remediation module, and an answer generation module. Each module can be implemented by software programs, hardware circuits, or a combination of software and hardware, and can be deployed in the same device or distributed across multiple computing nodes.

[0097] The knowledge graph construction module is used to acquire patent data to be processed and perform basic information processing and structured technical semantic extraction on it to obtain entity relationship data containing basic patent information, technical entities and technical relationships, and to construct an initial patent knowledge graph based on this.

[0098] The entity normalization and hybrid indexing module is used to normalize the technical entities based on vector similarity and patent classification number compatibility constraints, establish a mapping relationship from the original technical entities to the normalized technical entities, and construct a three-layer patent knowledge graph and hybrid retrieval index based on the initial patent knowledge graph and the mapping relationship, including a patent basic information layer, a technical semantic entity layer and a normalized entity layer.

[0099] The intent routing module is used to receive user natural language questions, perform intent recognition on the questions to generate a structured execution plan, and route the user natural language questions to a single patent question-and-answer track or a cross-patent analysis track accordingly.

[0100] The retrieval, verification, and remediation module is used to obtain candidate evidence and perform graph expansion based on the hybrid retrieval index in the routed track to obtain an evidence set; verify the evidence set; and perform at most one controlled remediation retrieval to update the evidence set when a remediable problem is detected.

[0101] The answer generation module is used to generate patent question-and-answer results based on the verified evidence set. The patent question-and-answer results include the answer content and the evidence path supporting the answer content.

[0102] Example 3: Electronic devices and computer-readable storage media

[0103] This invention also provides an electronic device, such as... Figure 7 As shown. The electronic device includes a processor and a memory, the memory storing a computer program, which, when executed by the processor, implements the steps in the above method embodiments.

[0104] The electronic device may include a server, desktop computer, laptop computer, cloud computing node, edge computing node, or other device with data processing capabilities. The processor may include a central processing unit, graphics processing unit, neural network processor, digital signal processor, or other processing unit capable of executing instructions. The memory may include read-only memory, random access memory, solid-state drive, hard disk drive, flash memory, or other media capable of storing computer programs and data.

[0105] This invention also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps in the above method embodiments. The computer-readable storage medium may be a non-transitory storage medium, such as a magnetic disk, optical disk, read-only memory, random access memory, flash memory, or solid-state memory.

[0106] The above embodiments are merely preferred embodiments of the present invention and are not intended to limit the present invention. Those skilled in the art can make various modifications, equivalent substitutions, and combinations to the above embodiments without departing from the spirit and substance of the present invention, and all such modifications, substitutions, and combinations should fall within the protection scope of the present invention.

Claims

1. A patent question-answering method based on knowledge graph-enhanced retrieval, characterized in that, Performed by an electronic device, including the following steps: S1. Obtain the patent data to be processed, perform basic information processing and structured technical semantic extraction on the patent data to be processed, obtain entity relationship data containing basic patent information, technical entities and technical relationships, and construct an initial patent knowledge graph based on it. S2. Based on vector similarity and patent classification number compatibility constraints, the technical entities under the same entity type are normalized to establish a mapping relationship from the original technical entities to the normalized technical entities. Based on the initial patent knowledge graph and the mapping relationship, a three-layer patent knowledge graph including a patent basic information layer, a technical semantic entity layer and a normalized entity layer is constructed, as well as a hybrid retrieval index including sparse index, dense vector index and relation path index. S3. Receive a user's natural language question, perform intent recognition on the question to generate a structured execution plan, and route the user's natural language question to a single patent question-and-answer track or a cross-patent analysis track accordingly. S4. In the routed track, candidate evidence is obtained based on the hybrid retrieval index and one-hop or two-hop graph expansion is performed to obtain an evidence set; the evidence set is verified for completeness, consistency and relevance, and at most one controlled remedy retrieval is performed to update the evidence set when a remedyable problem is detected. S5. Generate patent question-and-answer results based on the verified evidence set, wherein the patent question-and-answer results include the answer content and the evidence path supporting the answer content.

2. The method according to claim 1, characterized in that, In step S1, before performing structured technical semantic extraction, the publication number, title, abstract and patent metadata in the patent data to be processed are cleaned, split, deduplicated and standardized to obtain the basic patent information. The patent metadata includes at least one of the following: applicant, current right holder, inventor, IPC classification number, cited patent, application year, publication date and authorization status. The structured semantic extraction is performed through a structured semantic extraction model. The extraction results include a public number field, an entity array, and a relation array. Each entity in the entity array includes an entity name, an entity type, and an entity description. Each relation in the relation array includes a source node, a target node, a relation type, and a relation description. Before obtaining the entity relationship data, the extraction results are subjected to graph consistency verification, which includes enumeration consistency verification of entity type and relationship type, node existence verification of technical relationship, and isolated entity verification. If the verification fails, the extraction results are re-extracted or repaired at the field level.

3. The method according to claim 1, characterized in that, In step S2, the three-layer patent knowledge graph includes a patent basic information layer, a technical semantic entity layer, and a standardized entity layer; the patent basic information layer includes Patent nodes, Organization nodes, Person nodes, IPC nodes, and Year nodes; the technical semantic entity layer includes technical entity nodes, HAS_TECH_ENTITY relationships, and TECH_RELATION relationships; and the standardized entity layer includes CanonicalTechEntity nodes and MAPS_TO relationships. The standardization process includes: grouping the technical entities according to their entity types and embedding them into text; retrieving candidate synonyms or near-synonyms based on vector similarity; merging candidate entities with the same entity type, vector similarity meeting preset conditions, and IPC classification numbers of associated patents having at least one of the same major category, minor category, or main group into the same standardized technical entity; and establishing a mapping relationship between the candidate entities and the standardized technical entity.

4. The method according to claim 1, characterized in that, In step S3, the step of performing intent recognition on the question to generate a structured execution plan, and accordingly routing the user's natural language question to a single-patent question-and-answer track or a cross-patent analysis track, includes: Identify whether the user's natural language question contains publication numbers, patent titles, statistical terms, trend terms, or analytical terms; When a publication number or a uniquely identifiable patent title exists, the question type is determined to be a single patent Q&A, and the question is routed to the single patent Q&A track; When there are expressions related to quantity, proportion, statistics, or distribution, the question type is determined to be cross-patent statistical question and answer, and a corresponding execution identifier for cross-patent analysis track is generated; When there are expressions related to trends, evolution, hot topics, development characteristics, or technical routes, the question type is determined to be a cross-patent analysis question and answer, and a corresponding execution identifier for the cross-patent analysis track is generated; The core retrieval entities and graph expansion requirements are determined based on the identified technical entity words.

5. The method according to claim 1, characterized in that, In step S4, the hybrid retrieval index includes a keyword-based sparse index, a text embedding-based dense vector index, and a graph database-based relational path index. Candidate evidence is obtained based on the hybrid retrieval index, including: The search results are obtained through the sparse index and the dense vector index respectively, and then merged and sorted using the inverse sorting fusion method. For a candidate object d, its fusion score Score(d) is calculated based on the candidate object's ranking in different search result lists and a preset smoothing parameter, satisfying the following formula: , Where rankᵢ(d) is the ranking of candidate object d in the i-th search result list, and k is a preset smoothing parameter; And obtain the relationship path evidence associated with the candidate object through the relationship path index.

6. The method according to claim 1, characterized in that, In step S4: After being routed to the single-patent Q&A track, the target patent is preferentially matched based on the publication number or patent title; when the precise match fails, candidate patents are recalled based on the hybrid retrieval index, and the target patent is determined by sorting and fusion; starting from the Patent node corresponding to the target patent, a one-hop or two-hop local graph expansion is performed to obtain the technical problems, solutions, technical effects, core technical themes, industry fields and their relationship paths associated with the target patent, and these are added to the evidence set; After being routed to the cross-patent analysis track, multiple query variants are generated based on the user's natural language question. The search results are obtained using the hybrid search index based on the query variants. Seed patents, seed technology entities, or seed standard technology entities are determined by sorting and fusion. Based on the seed patents, seed technology entities, or seed standard technology entities, a graph expansion is performed to construct a topic-level semantic subgraph. The evidence set is added, and cross-patent statistics, topic summarization, trend analysis, or technology profiling analysis are performed based on the standardized entity layer.

7. The method according to claim 1, characterized in that, In step S4, the remediable problems include at least one of the following: empty technical chain, missing graph path, abnormal zero value statistics, off-topic candidate evidence, or insufficient answer evidence; the controlled remedial retrieval is executed at most once and includes at least one of the following: narrowing the keyword set, lowering the entity matching threshold, switching candidate seed entities, falling back to the original technical entity matching, falling back to the title or abstract text layer matching, and stopping the use of extended words that are prone to semantic drift. In step S5, the patent question-and-answer result is generated using structured evidence items as input. The structured evidence items include at least one of evidence source type, node identifier, node label, relationship type, relationship path, matching score, matching method, and supporting text. The patent question-and-answer result includes answer content, evidence path, and statistical caliber or uncertainty description corresponding to the question-and-answer track.

8. A patent question-answering system based on knowledge graph-enhanced retrieval, characterized in that, include: The knowledge graph construction module is used to acquire patent data to be processed, perform basic information processing and structured technical semantic extraction on the patent data to be processed, obtain entity relationship data containing basic patent information, technical entities and technical relationships, and construct an initial patent knowledge graph based on this. The entity normalization and hybrid indexing module is used to normalize the technical entities under the same entity type based on vector similarity and patent classification number compatibility constraints, establish the mapping relationship from the original technical entities to the normalized technical entities, and construct a three-layer patent knowledge graph including a patent basic information layer, a technical semantic entity layer and a normalized entity layer based on the initial patent knowledge graph and the mapping relationship, as well as a hybrid retrieval index including sparse index, dense vector index and relation path index. The intent routing module is used to receive user natural language questions, perform intent recognition on the questions to generate a structured execution plan, and route the user natural language questions to a single patent question-and-answer track or a cross-patent analysis track accordingly. The retrieval verification and remediation module is used to obtain candidate evidence based on the hybrid retrieval index and perform one-hop or two-hop graph expansion in the routed track to obtain an evidence set; to verify the completeness, consistency and relevance of the evidence set; and to perform at most one controlled remediation retrieval to update the evidence set when a remediable problem is detected. The answer generation module is used to generate patent question-and-answer results based on the verified evidence set. The patent question-and-answer results include the answer content and the evidence path supporting the answer content.

9. An electronic device, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program that, when executed by the processor, implements the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method according to any one of claims 1 to 7.

Images