Semantica 本体管理实战从知识图谱自动生成 OWL 本体并导出 Turtle / RDF/XML / JSON-LD【免费下载链接】semanticaGraph-Native Infrastructure for Context and Accountable AI Systems项目地址: https://gitcode.com/GitHub_Trending/sema/semantica导读OntologyGenerator是 Semantica 提供的本体自动生成引擎它直接从知识图谱中已有的实体与关系推导出一份正式的 OWL 本体无需预先设计 Schema随后可导出为 Turtle、OWL/XML、JSON-LD、N-Triples 等格式供 SHACL 约束校验、外部推理引擎HermiT、Pellet、ELK以及 STIX/TAXII 等工具链消费。读完本文你将掌握本体的核心概念、6 阶段生成流水线的底层原理、ClassInferrer/PropertyGenerator的增量扩展方法、LLM 引导式建本体的适用场景以及面向下游系统的完整导出与校验闭环。什么是本体Ontology本体是某个领域内概念与关系的形式化规范。它为知识图谱定义了一套共享词汇表核心由三部分构成类Classes——领域内实体的类型。以网络安全领域为例ThreatActor威胁行为体、Vulnerability漏洞、Software软件。类回答世界上存在哪些种类的事物。对象属性Object Properties——实体与实体之间的关系如exploits威胁行为体利用漏洞、targets恶意软件瞄准组织、uses行为体使用工具。数据类型属性Datatype Properties——携带字面量值的属性如name文本、severity_score小数、published_date日期、ip_address字符串。一个本体让知识图谱具备机器可读性它明确规定哪些关系是合法的、每个属性可以容纳什么类型的值、概念之间如何层级关联。这些约束正是自动化校验与推理的基础。为什么需要本体一致性Consistency没有本体时一个团队写Threat_Actor另一个团队写ThreatActor指代同一概念却产生两套命名。本体通过单一命名规范强制收敛。校验Validation本体支持自动检查——这条关系是否语义合理Vulnerability的 CVSS 分数应该存成文本还是数值推理Reasoning将本体导出为 OWL/Turtle 后外部推理引擎HermiT、Pellet、ELK可以推断出新事实。例如若Malware是Software的子类且HAMMERTOSS属于Malware则 OWL 推理器会推出HAMMERTOSS也属于Software。注意分工边界Semantica 负责导出本体推理本身在外部工具中运行。知识图谱质量Quality结构化 Schema 尽早暴露错误确保新数据与已有实体干净地集成。何时使用 / 何时不使用建议使用本体的场景实体关系复杂的正式知识图谱跨团队、跨组织的多源数据集成自动化推理与规则系统需要长期保持一致性的知识库需要对接期望 OWL/RDF Schema 的外部工具。不建议使用的场景面向文本的简单语义搜索向量相似度已足够的轻量 RAGSchema 快速变更的原型开发一次性数据分析不需要复用性本体的管理开销超过领域本身复杂度的情况。设计要点Semantica 的本体模块直接从知识图谱中已有的实体与关系推导正式 OWL 本体不需要预先设计 Schema。整个流程由 6 阶段流水线完成推断类、构建层级、映射 OWL 类型并序列化为 Turtle。流水线完全在内存中运行无需启动三元组存储——这一点在 Semantica 本体模块实现 的模块文档中有明确说明。一个简单示例从组织图谱到本体先用一个熟悉的业务域理解机制再进入网络安全场景。下面的代码直接构造一张小型组织图from semantica.ontology import OntologyGenerator # Build a simple organizational graph directly data { entities: [ {id: e-1, name: Alice, type: Person}, {id: e-2, name: Bob, type: Person}, {id: e-3, name: Carol, type: Person}, {id: e-4, name: Acme Corporation, type: Company}, {id: e-5, name: San Francisco, type: Location}, ], relationships: [ {source_id: e-1, target_id: e-4, type: works_for}, {source_id: e-2, target_id: e-4, type: works_for}, {source_id: e-1, target_id: e-3, type: reports_to}, {source_id: e-4, target_id: e-5, type: headquartered_in}, ], } generator OntologyGenerator( base_urihttps://company.example.org/ontology/, min_occurrences1, ) ontology generator.generate_ontology( data, nameOrganizationOntology, build_hierarchyTrue, ) # Inspect what was generated print(fClasses: {len(ontology.get(classes, []))}) # Classes: 3 print(Object Properties:) for prop in ontology.get(properties, []): if prop.get(type) object: domain , .join(prop.get(domain, [])) range_ , .join(prop.get(range, [])) print(f {prop[name]} ({domain} → {range_})) # works_for (Person → Company) # reports_to (Person → Person) # headquartered_in (Company → Location) print(Datatype Properties:) for prop in ontology.get(properties, []): if prop.get(type) data: print(f {prop[name]} ({prop.get(range)})) # name (string)关键参数解读base_uri如https://company.example.org/ontology/会成为所有类与属性的命名空间前缀。Person在导出的 RDF 中表现为https://company.example.org/ontology/Person。从源码看类名由 NamespaceManager.generate_class_iri 统一转换为PascalCase后拼接到 base URI 上当use_speaking_iris开启默认开启时生成可读 IRI同时内部注册了 rdf、rdfs、owl、xsd、skos、dc、dcterms 等标准命名空间。min_occurrences实体类型出现多少次才升级为类。示例传1表示每种出现过的实体类型都成为类。注意 ClassInferrer 的默认值是 2——若想对罕见类型也建类需要显式调低。对象属性连接实体与实体works_for、reports_to其 domain/range 由关系两端的实体类型推断数据类型属性连接实体与字面量name。一张没有 Schema 的图有了 Schema 的概念下面用 CTI网络威胁情报数据填充知识图谱让流水线机制可见。此时图谱里已经有一批节点与边但没有任何正式 SchemaAPT29和假设中的Lazarus Group都是 ThreatActor但没有任何机制强制二者都必须携带attribution_confidence属性。from semantica.context import AgentContext, ContextGraph from semantica.vector_store import VectorStore vs VectorStore(backendfaiss, dimension768) graph ContextGraph() ctx AgentContext(vector_storevs, knowledge_graphgraph, graph_expansionTrue) ctx.store( [ CVE-2024-3400 is a critical vulnerability in PAN-OS exploited by APT29., APT29 is a Russian state-sponsored threat actor targeting NATO governments., PAN-OS is a network operating system developed by Palo Alto Networks., HAMMERTOSS is a backdoor malware used by APT29 for command-and-control., ], extract_entitiesTrue, extract_relationshipsTrue, ) # At this point we have ~8 nodes and several edges, but no formal schema. print(fGraph nodes: {len(graph.to_dict().get(nodes, []))})这里用到的AgentContext/ContextGraph/VectorStore构成了本体生成的数据源上下文图。文本经extract_entitiesTrue与extract_relationshipsTrue完成实体与关系抽取该抽取能力见 语义抽取指南图谱随后即可作为本体的输入。生成本体6 阶段流水线OntologyGenerator读取图谱 dict 并运行完整流水线。generate_from_graph是generate_ontology的别名见 ontology_generator.py 实现入参即为graph.to_dict()的输出。from semantica.ontology import OntologyGenerator generator OntologyGenerator( base_urihttps://cti.example.org/ontology/, min_occurrences1, # every entity type that appears at least once becomes a class ) ontology generator.generate_from_graph( graph.to_dict(), nameCyberThreatOntology, build_hierarchyTrue, # infer parent-child class relationships ) # What the pipeline produced: print(fClasses : {len(ontology.get(classes, []))}) # Classes : 4 → ThreatActor, Vulnerability, Software, Malware print(fProperties: {len(ontology.get(properties, []))}) # Properties: 3 → exploits (object), targets (object), name (datatype) # Inspect a class: for cls in ontology.get(classes, []): print(f {cls[name]} parent{cls.get(parent)}) # ThreatActor parentNone # Vulnerability parentNone # Software parentNone # Malware parentSoftware ← hierarchy inferred because HAMMERTOSS was linked to PAN-OS (Software)流水线的 6 个阶段源码级拆解从 OntologyGenerator 类文档 与 模块总览 可以完整还原 6 个阶段的职责阶段名称职责Stage 1Semantic Network Parsing从 entities / relationships 提取领域概念实体类型计数Counter、关系模式抽取、概念分组同时兼容Entity对象、[text, label]列表、RDF 三元组[source, type, target]等多种输入形态Stage 2YAML-to-Definition把概念转换为类定义结构调用ClassInferrer.infer_classes()Stage 3Definition-to-Types映射 OWL 类型为每个类标注type: owl:Class对象属性标注owl:ObjectProperty数据属性标注owl:DatatypeProperty调用PropertyGenerator.infer_properties()Stage 4Hierarchy Generation构建分类学结构父子关系推断、传递闭包计算、DFS 环形依赖检测Stage 5TTL Generation使用 rdflib 生成 OWL/Turtle 语法处理命名空间前缀与三元组序列化由 OWLGenerator 承接Stage 6Symbolic ValidationHermiT / Pellet 推理校验一致性consistency、可满足性satisfiability与结构校验结果写入ontology[validation]两个值得注意的实现细节类的 URI 兜底Stage 3 中若类没有uri会通过namespace_manager.generate_class_iri()补齐对象属性同理使用generate_property_iri()属性名为 camelCase。这是历史上修复类缺失 IRI 导致导出失败问题issue #1103的关键逻辑。校验默认开启options.get(validate, True)意味着生成即校验校验结果valid/consistent/satisfiable/errors/warnings直接附加在返回的 ontology dict 上见 ontology_generator.py 的 Stage 6 代码。Malware的层级条目说明流水线基于关系图中的共现模式检测到恶意软件是软件的子类型因为 HAMMERTOSS 与 PAN-OS/Software 建立了关联。ClassInferrer.build_class_hierarchy 还包含两条启发式按命名拆分找更一般的父类如Manager→Employee以及在Entity/Thing/Resource等常见父类中匹配。这些自动推断在导出前可以手工覆盖。校验 Schema导出前先做结构校验from semantica.ontology import validate_ontology result validate_ontology(ontology) print(fValid: {result.get(valid, False)}) # Valid: True for w in result.get(warnings, []): print(f WARN : {w}) # WARN : Class Malware has no declared datatype properties for e in result.get(errors, []): print(f ERROR: {e}) # (none — the ontology is structurally sound)类缺少数据类型属性的警告在此阶段很常见它表示推理流水线在图谱中发现了该类但没有任何节点携带显式属性值。可以在下一轮导出前用ClassInferrer与PropertyGenerator手工补充属性。校验层的完整能力不止于结构检查OntologyValidator还支持一致性consistency与可满足性satisfiability检查并提供了 SHACL 相关模型SHACLValidationReport、SHACLViolation与run_shacl_validation()——后者封装 pySHACL 校验并生成可读的违规解释依赖通过pip install semantica[shacl]安装详见 ontology_validator.py。增量扩展新实体类型出现时如何成长用ClassInferrer逐批新增实体类型无需重新生成整个本体from semantica.ontology import ClassInferrer, PropertyGenerator # Fine-grained control: infer classes from the new batch of entities new_entities [ {id: kev-001, name: KEV-CVE-2024-3400, type: KEVEntry, due_date: 2024-04-19, ransomware_use: Known}, {id: kev-002, name: KEV-CVE-2023-4966, type: KEVEntry, due_date: 2023-11-14, ransomware_use: Unknown}, ] inferrer ClassInferrer() new_classes inferrer.infer_classes(new_entities) # [{id: KEVEntry, name: KEVEntry, parent: None, ...}] # Manually set the parent to Vulnerability before merging for cls in new_classes: if cls[name] KEVEntry: cls[parent] https://cti.example.org/ontology/Vulnerability # Inspect the hierarchy hierarchy inferrer.build_class_hierarchy(new_classes) print(hierarchy) # {KEVEntry: {parent: ...Vulnerability, children: []}} # Merge into the existing ontology ontology[classes].extend(new_classes) # Infer properties from the new nodes prop_gen PropertyGenerator() # PropertyGenerator reads entity attributes and relationship patterns # to produce datatype properties (due_date: xsd:date) and object properties这里的模式是增量式的对每个新批次运行infer_classes人工审查结果在流水线猜错父类时手动调整parent赋值然后合并。本体随图谱一起成长而不是落后于图谱。底层行为值得注意源码依据类名规范化冲突检测infer_classes会先对实体类型做命名规范化若多个原始类型归一化后产生重复类名会抛出ValidationError见 class_inferrer.py 碰撞检测。公共属性抽取阈值某属性只有在 ≥50% 的同类实体中出现才会被列为该类的公共属性_extract_common_properties。属性名冲突检测PropertyGenerator会合并规范化后同名的属性若对象属性与数据属性撞名则抛出ValidationError见 property_generator.py 的_coalesce_normalized_properties。XSD 类型自动推断PropertyGenerator._infer_property_type依据 Python 值类型映射——bool → xsd:boolean、int → xsd:integer、float → xsd:double、形如YYYY-MM-DD的字符串 →xsd:date、含时间戳的字符串 →xsd:dateTime其余 →xsd:string多类型冲突时按类型层级boolean integer double date dateTime string取更一般者源码位置。示例中的due_date: 2024-04-19因此被推断为xsd:date。元字段白名单id、type、entity_type、text、label、confidence、properties、relationships、metadata、merged_from、merge_strategy等控制字段不会被误当成数据类型属性_CONTROL_FIELDS源码。从非结构化文本生成本体当还没有结构化图谱时LLMOntologyGenerator可以借助 LLM 从散文直接提取类与属性适合领域引导bootstrappingfrom semantica.ontology import LLMOntologyGenerator llm_gen LLMOntologyGenerator(providergroq, modelllama-3.1-8b-instant) ontology_from_text llm_gen.generate_ontology_from_text( APT29 (also known as Cozy Bear) is a Russian state-sponsored threat actor. They use spear-phishing emails to deliver HAMMERTOSS malware. HAMMERTOSS communicates over Twitter and GitHub to evade detection. The group has been observed exploiting CVE-2024-3400 in PAN-OS appliances. ) # The LLM identified: ThreatActor, Malware, Vulnerability, Platform, CommunicationChannel print(fClasses extracted: {len(ontology_from_text.get(classes, []))}) # Supported providers: groq, openai, anthropic, novita从 llm_generator.py 源码 可见其实现机制provider默认openai通过semantic_extract.providers.create_provider()创建 LLM Provider并调用generate_structured()让模型按固定 JSON Schema 输出类与属性随后_normalize_output()统一补全uri、label、comment、parent等字段URI 缺失时以base_uri拼接类名/属性名兜底metadata中记录source: llm与 provider 名称。使用建议LLMOntologyGenerator最适合在没有结构化图谱的新领域做冷启动。一旦有了图谱优先使用OntologyGenerator.generate_from_graph()——它是确定性的、可复现的且每次运行不消耗 LLM token。面向下游系统导出按下游工具期望的格式导出from semantica.export import export_owl, export_rdf # OWL/XML — for Protégé, OWL API, HermiT, Pellet export_owl(ontology, cyber_threat.owl, formatowl-xml) # Turtle — compact, human-readable; preferred for SHACL toolchains export_rdf(ontology, cyber_threat.ttl, formatturtle) # JSON-LD — for web APIs and linked-data applications export_rdf(ontology, cyber_threat.jsonld, formatjsonld) # N-Triples — for bulk load into triple stores (GraphDB, Stardog, Oxigraph) export_rdf(ontology, cyber_threat.nt, formatntriples)两个导出函数的完整能力export/methods.py 源码export_rdf支持turtle、rdfxml、jsonld、ntriples、n3五种格式内部由RDFExporter完成序列化定义位置。export_owl支持owl-xml默认与turtle两种格式内部由OWLExporter完成定义位置。示例中传给formatturtle的用法是受支持的。导出的 Turtle 文件正是 Semantica SHACL 校验管线的输入。本体不仅用于导出SHACLGenerator位于 ontology_generator.py可以从本体 dict 生成 W3C SHACL 约束形状支持 Turtle / JSON-LD / N-Triples 三种序列化提供standard/strict质量档位strict档会对所有声明了属性的形状启用sh:closed并支持父类形状向子类传播include_inherited。如何在运行时生成约束形状并校验真实图数据见 SHACL 校验指南。常见陷阱过度建模Over-modeling。不要造 50 个类而实际 10 个就够。从简单开始只有在推理或校验确实需要形式化区分时才增加复杂度。比如MaliciousEmail与PhishingEmail分开建类只有当二者拥有不同属性或关系时才有意义。本体漂移Ontology drift。图谱中持续出现新实体类型时若不重新生成或增量更新本体会逐渐过时。应建立监控检测当前本体未覆盖的新实体类型出现。类命名不一致。选定一种约定CamelCase、snake_case 或 kebab-case并坚持使用。在同一本体中混用ThreatActor、threat_actor、threat-actor会造成混乱并破坏期望统一命名的下游工具。Semantica 自身的 NamingConventions 模块会强制类名 PascalCase、属性名 camelCase并在校验时给出改名建议。忽略生成即校验。OntologyGenerator默认会在 Stage 6 运行校验并把validation结果挂到返回 dict 上——建议在导出前主动检查ontology[validation][valid]避免把有结构问题的本体发布给下游。领域示例场景一Defense — CTI / 威胁情报防御型 CTI 团队每天早上摄入原始 OSINT 报告。本体必须与 STIX 2.1 和 NATO MISP 分类法保持互操作因此 IRI 遵循 DoD 命名空间本体导出为 OWL/XML 供组织内 SIEM 推理插件使用from semantica.context import AgentContext, ContextGraph from semantica.vector_store import VectorStore from semantica.ingest import ingest_file, ingest_web from semantica.ontology import OntologyGenerator, validate_ontology from semantica.export import export_owl, export_rdf vs VectorStore(backendfaiss, dimension768) graph ContextGraph() ctx AgentContext(vector_storevs, knowledge_graphgraph, graph_expansionTrue) # Ingest a campaign PDF and NVD advisory page cti_report ingest_file(apt29_cozycar_2024.pdf, methodfile) nvd_entry ingest_web(https://nvd.nist.gov/vuln/detail/CVE-2024-3400, methodurl) ctx.store( [cti_report.text, nvd_entry.text], extract_entitiesTrue, extract_relationshipsTrue, ) generator OntologyGenerator( base_urihttps://ontology.dod.mil/cyber/, min_occurrences1, ) ontology generator.generate_from_graph( graph.to_dict(), nameCyberThreatOntology, build_hierarchyTrue, ) result validate_ontology(ontology) print(fOntology valid: {result.get(valid)}) # Ontology valid: True # OWL/XML for the SIEM reasoning plugin; Turtle for the SHACL pipeline export_owl(ontology, ./ontologies/cyber_threat.owl, formatowl-xml) export_rdf(ontology, ./ontologies/cyber_threat.ttl, formatturtle) print(fClasses : {len(ontology.get(classes, []))}) print(fProperties : {len(ontology.get(properties, []))}) # Classes : 7 (ThreatActor, Vulnerability, Malware, Platform, Campaign, ...) # Properties : 9 (exploits, targets, uses, name, cvss_score, ...)场景二Security — SOC / 安全运营中心SOC 团队将零信任身份实体——用户、服务账号、资源与策略——建模为 OWL 本体让策略评估引擎使用共享的形式化词汇表而不是硬编码字符串from semantica.ontology import OntologyGenerator, ClassInferrer from semantica.export import export_owl # Hand-specify the identity graph — in production this comes from your IAM export data { entities: [ {id: u-1, name: alice, type: User}, {id: u-2, name: svc-scanner,type: ServiceAccount}, {id: r-1, name: kube-api, type: Resource}, {id: r-2, name: s3-prod, type: Resource}, {id: p-1, name: ReadOnly, type: Policy}, {id: p-2, name: AdminAccess,type: Policy}, ], relationships: [ {source_id: u-1, target_id: p-1, type: BOUND_TO}, {source_id: u-2, target_id: p-2, type: BOUND_TO}, {source_id: p-1, target_id: r-1, type: ALLOWS_ACCESS}, {source_id: p-2, target_id: r-2, type: ALLOWS_ACCESS}, ], } generator OntologyGenerator( base_urihttps://zerotrust.corp/ontology/, min_occurrences1, ) ontology generator.generate_ontology(data, nameZeroTrustOntology) # Inspect what the pipeline inferred inferrer ClassInferrer() classes inferrer.infer_classes(data[entities]) for cls in classes: print(f Class: {cls.get(name)}) # Class: User # Class: ServiceAccount # Class: Resource # Class: Policy export_owl(ontology, ./ontologies/zero_trust.owl, formatowl-xml) print(Ontology exported for policy evaluation engine)场景三Life Science — 临床 / 制药制药团队需要一份对齐 OBO Foundry 约定GO、CHEBI、HP的本体来描述 II/III 期肿瘤试验方案。由于源数据保存在 PostgreSQL 试验数据库而非知识图谱中他们用LLMOntologyGenerator从散文描述引导建本体from semantica.ontology import LLMOntologyGenerator, OntologyGenerator, validate_ontology from semantica.export import export_owl, export_rdf from semantica.ingest import DBIngestor # Load trial records from the clinical database db DBIngestor() trial_rows db.execute_query( postgresql://readonlyclindb:5432/trials, SELECT compound, target_protein, disease_indication, mechanism_of_action, primary_endpoint FROM trial_protocols WHERE phase IN (II,III) , ) # Construct a natural-language summary for the LLM protocol_text \n.join( fCompound {r[compound]} targets {r[target_protein]} fin {r[disease_indication]} via {r[mechanism_of_action]}. fPrimary endpoint: {r[primary_endpoint]}. for r in trial_rows ) # LLM extracts classes: Compound, TargetProtein, DiseaseIndication, # MechanismOfAction, ClinicalEndpoint, ClinicalTrial llm_gen LLMOntologyGenerator(provideropenai, modelgpt-4o) ontology llm_gen.generate_ontology_from_text(protocol_text) result validate_ontology(ontology) print(fValid: {result.get(valid)}) for w in result.get(warnings, []): print(f WARN: {w}) # Export aligned to OBO Foundry URI convention export_owl(ontology, ./ontologies/clinical_trial.owl, formatowl-xml) export_rdf(ontology, ./ontologies/clinical_trial.ttl, formatturtle) print(Ontology ready for Protégé review and OBO alignment check)场景四Banking — 风险 / 合规风险团队将 Basel III / BCBS 239 概念形式化为 OWL 本体让自动化合规规则能够以共享词汇表对信用风险实体进行推理——取代 Python 脚本中硬编码的字段名检查from semantica.ontology import LLMOntologyGenerator, validate_ontology from semantica.export import export_owl, export_rdf from semantica.ingest import ingest_file # Ingest regulatory source documents regs [ ingest_file(basel3_cre20.pdf, methodfile), ingest_file(sr_11_7.pdf, methodfile), ingest_file(bcbs239.pdf, methodfile), ] # Use an LLM to extract the conceptual model from regulatory prose llm_gen LLMOntologyGenerator(provideranthropic, modelclaude-sonnet-5) ontology llm_gen.generate_ontology_from_text( \n\n.join(r.text[:8000] for r in regs) # token-safe excerpt per document ) result validate_ontology(ontology) if not result.get(valid): for err in result.get(errors, []): print(fERROR: {err}) # Fix errors before publishing to the compliance rule engine else: print(Ontology valid — publishing to compliance registry) # Turtle for SHACL shapes; OWL/XML for HermiT reasoning; JSON-LD for the API export_owl(ontology, ./ontologies/regulatory.owl, formatowl-xml) export_rdf(ontology, ./ontologies/regulatory.ttl, formatturtle) export_rdf(ontology, ./ontologies/regulatory.jsonld, formatjsonld)配置方式补充除了在构造函数中传参本体模块还支持通过 OntologyConfig 集中管理配置优先级为配置文件 → 环境变量 → 默认值环境变量ONTOLOGY_BASE_URI、ONTOLOGY_MIN_OCCURRENCES、ONTOLOGY_REASONER、ONTOLOGY_FORMAT、ONTOLOGY_SIMILARITY_THRESHOLD、ONTOLOGY_CHECK_CONSISTENCY、ONTOLOGY_CHECK_SATISFIABILITY等源码映射表配置文件YAML / JSON / TOML 均可读取其中ontology与ontology_methods两个顶层键程序化配置ontology_config.set(base_uri, https://example.org/ontology/)或set_method_config(generate, min_occurrences1)。相关指南SHACL 校验——从本体生成 W3C SHACL 约束形状并用其对实时图数据做校验推理与规则——对本体应用前向/后向链式规则以推导新事实导出与序列化——将图谱导出为 RDF、GraphML、CSV 与 Neo4j Cypher语义抽取——抽取喂给本体生成的实体与关系上下文图谱——本体生成所读取的知识图谱【免费下载链接】semanticaGraph-Native Infrastructure for Context and Accountable AI Systems项目地址: https://gitcode.com/GitHub_Trending/sema/semantica创作声明：本文部分内容由AI辅助生成（AIGC），仅供参考