RESEARCH DOSSIER
Legal Knowledge Graphs
Abstract. Legal knowledge graphs encode legal documents, concepts, institutions, parties, authorities, and relationships as structured, queryable assertions. Their scholarly usefulness depends on whether the underlying assertions preserve the distinctions on which legal meaning turns: the identity of an enacted instrument and its versions; the difference between textual citation and legal reliance; the relation between a provision's publication, commencement, applicability, amendment, and repeal; and the difference between an extraction result and an authenticated legal proposition. This monograph examines graph representation in the light of Semantic Web standards, legal ontologies, official European legislation and case-law identifiers, structured legislative documents, provenance standards, and current research on legal text processing. It develops a source-centered analytical account of how such graphs can support legal research while remaining accountable to official records and interpretive context. The inquiry combines doctrinal source analysis with technical comparison and expressly hypothetical examples. Its conclusions distinguish machine-verifiable graph syntax and shape constraints from juridical validity, distinguish reference relationships from normative authority, and show why temporal precision, source attribution, jurisdiction and correction procedures must be built into every consequential legal graph query. No new proprietary ontology, named theoretical framework, or empirical performance claim is proposed.
Keywords: Legal knowledge graphs; legal informatics; knowledge representation; RDF; OWL; SPARQL; SHACL; legal ontology; ELI; ECLI; Akoma Ntoso; LegalRuleML; legal citation; legal provenance; temporal law; entity resolution; graph validation; comparative legal research
Research Problem and Definition
A legal knowledge graph is a network of identified entities and asserted relations designed to represent information about law in a form that computers can index, traverse, validate, and query. Its nodes may identify an act, an article, an amendment, a judicial decision, a court, a legal concept, a territorial jurisdiction, a date, or a particular textual expression. Its edges may record that one decision cites another, that a provision is contained in an instrument, that an amending act changes a specified paragraph, or that a research classification associates two subjects. Each edge is an assertion attributable to a source or processing activity. The legal researcher must be able to determine what the assertion means, who made it, and whether it is supported.
This definition encompasses RDF graphs, labeled property graphs, and other typed relational representations, while emphasizing the form and epistemic function of the relationships rather than an exclusive storage technology. The RDF 1.1 data model defines a graph as a set of subject–predicate–object triples and introduces datasets containing a default graph and named graphs. Legal graph practice adds a demanding interpretive problem: a triple identifying a citation is often objectively recoverable from the text, whereas a triple asserting that a proposition is legally binding can require jurisdiction-specific doctrine, interpretation, and assessment of subsequent authority.
The core research question is how legal norms, judicial decisions, institutions, and their changing relations can be represented as auditable structures without losing the distinctions that make a legal assertion authoritative or contestable. A useful graph can identify relations between sources and expose the route by which a researcher reached a conclusion. It can also accumulate misleading certainty when document metadata, extractor predictions, and legal conclusions share the same unqualified predicate. The quality of a legal graph therefore rests on data modeling, bibliographic discipline, and jurisprudential analysis at once.
A graph is a means of organizing propositions about law. An authoritative legislature, court, or publisher remains the source of the legal act or decision; the graph supplies a structured representation with its own evidentiary and computational properties. The European Legislation Identifier, for example, makes legislative information addressable through standardized identifiers and metadata. An ELI reference can identify an instrument or version. Whether a disputed rule governs a particular event requires the text, jurisdiction, dates and interpretive authorities to be examined.
Legal Information, Knowledge Representation, and Research Traditions
Legal knowledge representation developed through several partly overlapping traditions: formal logic and rule modeling; legal information retrieval and document indexing; ontologies of legal concepts; citation network analysis; and linked-data publication. Their distinct purposes explain why the same document can be represented differently. A full-text index optimizes retrieval by language. A citation network captures a document-to-document relation. A norm model attempts to express conditions and consequences. An institutional graph locates a text among courts, legislatures and agencies. A single data platform may combine all four, but their assertions must remain distinguishable.
The legal ontology tradition investigates the categories through which legal meanings are structured. Hoekstra, Breuker, Di Bello, and Boer (2007) developed the LKIF Core ontology as a repertoire of basic legal concepts, presenting its methodology and use in formalizing an EU directive. Its scholarly significance lies in making conceptual commitments inspectable and reusable. The ontology is an analytic instrument, while the source of law and the defensibility of an interpretation are separate questions.
Another tradition concentrates on the texts of legislation, debate and adjudication. The Akoma Ntoso XML Vocabulary, standardized by OASIS in 2018, offers structured elements and metadata for legal documents. Structured documents establish boundaries for acts, sections and references that graph builders can use as evidence. A machine-readable document itself is not yet a full graph of legal meaning; it is an unusually valuable ground truth for identifying the text and its internal relations.
Legal information retrieval supplies additional methodological resources. Ashley's (2017) Artificial Intelligence and Legal Analytics situates computational techniques within the practical tasks of legal reasoning and search. It provides a conceptual literature anchor rather than proof that any individual legal graph automatically improves research outcomes. Legal NLP surveys identify document length, complex terminology, limited open data, and evaluation challenges as recurrent practical issues. Ariai and Demartini (2024) discuss these constraints in their survey of legal language processing.
The resulting historical picture is plural. Formal rules, explicit ontologies, searchable case texts, and graph relations solve different portions of the research problem. An effective legal knowledge graph makes these distinctions visible in its data rather than treating a shared storage format as evidence of a shared level of meaning.
Sources, Method, and Standards of Inference
The inquiry uses documentary analysis of published technical standards; doctrinal interpretation of official legal-source and identifier guidance; comparison of institutional implementations; and critical reading of selected academic literature. The primary evidence includes W3C Recommendations on RDF, OWL, SPARQL, JSON-LD, SHACL and provenance; OASIS standards for legal XML and rule interchange; and official European and British legal publication documentation. Academic literature is used for identified methodological positions and research context. Dates and formal standards status are recorded as of November 21, 2026.
Source classes carry different evidentiary weight for different claims. An enacted statute or court judgment is evidence of a legal source within its institutional and temporal setting. A publisher's explanatory page is evidence of how its platform identifies and supplies documents. A W3C Recommendation is evidence of the defined computational language and conformance requirements. An academic paper is evidence of its authors' method, findings and arguments to the extent actually available; its abstract cannot prove unstated results. This allocation of evidentiary roles governs every chapter.
Methodologically, the work distinguishes five operations: identification of a source object; formal representation of a sourced assertion; validation of data constraints; legal interpretation of the assertion; and possible empirical evaluation of systems using the graph. These are analytical distinctions within established research practice, rather than a new theory or prescribed universal architecture. They prevent a valid RDF triple or a successful extraction benchmark from being treated as equivalent to an established legal conclusion.
The comparison emphasizes European public-law data because official documentation for ELI, ECLI and the Publications Office's Cellar provides concrete, verifiable implementation evidence. United Kingdom legislation documentation supplies a complementary example of revision and change tracking. These contexts are used to analyze graph design questions, not to announce the existence of uniform legislation across Europe or to attribute the same binding effects to different national systems.
Illustrative triples, identifiers, temporal situations and query scenarios in this monograph are conceptual examples. They were not extracted from a deployed graph, benchmarked against a live endpoint, or validated against a real case unless an expressly cited official example says so. Source bibliographies are recorded in References and a separate machine-readable evidence register with the actual research access level for each item.
Graph Models, Data Structures, and Semantic Commitments
In RDF, a node is typically identified by an IRI or represented as a blank node; an object can also be a literal. The three positions in a triple make a statement about a relation between resources. A legal example could assert that an identified statutory section is part of an identified act. A second assertion could record that a court opinion mentions that section. A third could state that an official repository publishes a particular language expression. These relations are semantically different and should have separate predicates.
The SPARQL 1.1 specification provides a query language for graph patterns, optional matches, filters, aggregation, and named datasets. It permits a researcher to ask which instruments amend a given act, or which decisions cite a specified authority, provided the underlying data contains the relevant assertions. Results report what matches the modeled dataset and inference regime. They do not establish that every real-world authority has been indexed. A query returning no matches may indicate a negative answer in a closed and complete local register, or merely absent coverage in an open and incomplete corpus.
OWL adds formally specified vocabulary and inference capabilities. The W3C OWL 2 document overview describes classes, properties, individuals, and data values with explicit semantics. In legal modeling, an OWL class such as JudicialDecision can help describe source types; property axioms can support consistent navigation. Yet categorical legal questions, such as whether a tribunal is bound by another court in a certain procedural posture, demand precise source-based rules. Broad axioms should not silently transform a local precedent relation into a universal transitive hierarchy.
Property graphs often attach properties directly to nodes and edges and can serve investigative, operational, and network-analytics workflows. RDF environments prioritize shared identifiers, published vocabularies, formal semantics and interoperable exchange. The decision between them is partly practical: available source data, cross-institution linkage, latency, permissions, and query requirements matter. Both can preserve source-specific provenance if the model is designed to do so. Neither storage family independently determines legal truth.
The JSON-LD 1.1 Recommendation makes linked data expressible through a JSON-based serialization. This is useful for publishing legal metadata on the Web and transferring data between systems, especially when the metadata vocabulary and IRIs are documented. Serialization choice should be understood as an exchange mechanism, with graph meaning and provenance defined at a separate level.
Legal Ontologies and the Limits of Concept Alignment
A legal ontology declares the types of things and relations that a graph is prepared to recognize. Familiar distinctions include normative instrument, source expression, jurisdiction, institution, procedural role, legal person, sanction, obligation, permission, and adjudicative event. The most consequential choices involve apparently common words whose meanings vary by legal system. A term such as regulation, court, charge, consideration, equity, or possession can classify different legal objects in distinct jurisdictions and periods. The exact same text label should therefore be insufficient evidence of conceptual identity.
The LKIF Core study shows why explicitly articulated legal concepts are valuable for representation and reuse. An ontology can declare that a model's concept has a particular definition and relation to other concepts. It still requires an account of the authority for applying that category to a given text, and of the translator or analyst responsible for a cross-system mapping. Conceptual reuse is credible when the target domain's conditions are examined rather than assumed.
An ontology may be highly formal while the mapping of one jurisdiction's legal category to another remains interpretive. Exact equivalence claims should be reserved for genuinely coextensive terms. Broader or narrower relations, associated subjects and translation correspondences are often the more accurate representation. The W3C SKOS Reference explicitly supports conceptual schemes, multilingual labels, hierarchical and associative links, and mapping relations. Its lighter vocabulary can be preferable for catalog navigation when a strong ontological identity claim is unwarranted.
Legal datasets also require a distinction between what a source calls itself and what a researcher infers it to be. A ministry may label an instrument an order; a comparative classification system may group it with delegated legislation. Both statements can be recorded if source attribution and conceptual perspective are retained. A single unqualified classification would erase the provenance of the competing descriptions. Ontology alignment should expose uncertainty and review status, particularly for historical law and multilingual material.
Identifiers, Entities, and Referential Integrity
Identity is the first condition for meaningful legal linking. An identifier may refer to a legislative work, a particular version, a language expression, a publication file, a cited provision, or a judiciary's decision record. Confusing those levels produces errors that remain syntactically invisible. A court's ECLI identifies a decision within a defined identifier scheme; the PDF and HTML files carrying its text are manifestations or representations supplied by repositories. A legislative identifier may support separate paths for acts and their subdivisions.
The ELI specification guidance explains the use of HTTP URIs to identify legislation and the metadata needed for exchange. The official documentation illustrates a URI for a specific article, showing why provision-level linking is possible where the publisher implements it. The ECLI identifier guidance describes a five-part decision identifier that includes jurisdiction, court, year, and a distinguishing number. These schemes solve different identity problems and should not be treated as interchangeable IDs.
An institution's administrative name is likewise distinct from the identity of its legal powers through time. Courts can be renamed or reorganized; statutory divisions can be renumbered; publishers can change document URLs; cases can have multiple reported citations. A robust graph retains stable subject identifiers when possible, records alternative labels and citations as properties, and links successor entities explicitly rather than silently equating them.
A graph's entity-resolution pipeline should retain the original mention and the evidence for attaching it to a canonical node. For example, a case extraction might identify a short-form citation that plausibly refers to two decisions. The safe representation records the candidate matches and review status until the deciding context is found. Premature conflation can manufacture citation paths, contaminate retrieval and carry the mistake into later computational analyses.
Referential integrity includes periodic monitoring. A stable URI policy addresses naming and redirection; content hashes identify fixed documentary bytes; version records identify the edition to which an assertion relates. These safeguards serve different purposes and can be combined without treating a hash as proof of legal authenticity or a working hyperlink as proof of doctrinal currency.
Legislative Documents as Structured Legal Objects
Legislation has an internal documentary anatomy that determines the granularity of a useful graph. The instrument has a title, enactment history, publishing authority, document type and possibly multiple language expressions. Within it lie articles, sections, schedules, clauses, definitions, headings, tables and references. Different jurisdictions assign different legal effects to these elements. A graph designed for clause-level search should therefore identify subdivisions according to the authoritative text's own hierarchy and preserve the text surrounding each asserted relationship.
The Akoma Ntoso document vocabulary supplies an XML-oriented representation of legislative, parliamentary and judicial documents, including structural and metadata elements. It can carry an identifiable documentary hierarchy and references from which graph assertions may be derived. By contrast, a semantic graph may record that two sections have a thematic or legal relation even when the connection is not explicitly encoded in the XML. A pipeline should preserve the difference between a structural fact extracted from a document element and a proposition constructed through interpretation or language processing.
Citation to a provision ought to point to the smallest stable unit that can carry the proposition. A whole-act link is insufficient when a legal argument turns on a definition in one section. Conversely, citing a subordinate clause without its exceptions, commencement information, and applicable schedules can distort the rule. The appropriate unit of retrieval depends on the legal task and should support reassembly of the surrounding instrument.
Publishers may issue corrigenda, replacing or correcting published text without always using the same legislative technique as an amendment. A good model therefore distinguishes the textual event, the legal event, and the repository event. A corrected digital file should not be mistaken for a new legislature's enactment; an amendment should not be mistaken for a mere metadata correction. Where an official record describes these acts, the graph can attach them to the affected instrument and version using separate, documented relations.
Document structure is the foundation of source-specific extraction, but structure does not complete legal interpretation. The decisive scholarly requirement is an auditable passage from the legal text to the graph assertion, with links back to official words, a documented version, and any interpretive step used to produce a higher-level claim.
Cases, Citations, and Judicial Relationships
Judicial records form a different graph from legislation. A decision is an event and a text issued by a court in a particular procedural setting. It has parties, dates, an outcome, possibly a panel and separate opinions, and relations to earlier cases and enacted authorities. The presence of a citation establishes that the later text refers to a named source; further doctrinal interpretation is needed to say whether the court followed, distinguished, limited, doubted, or overruled a proposition.
The European ECLI metadata framework standardizes a decision identifier and a minimum set of metadata for cross-system discovery. It facilitates linking judgments without solving the jurisprudential analysis of the relationship between them. The Court of Justice citation guidance illustrates the use of ECLI to make European judicial citations consistent and readily traceable.
A citation graph is best viewed as a primary documentary layer supplemented by carefully attributed interpretive classifications. One edge can record that Opinion A cites Opinion B, with the passage showing where. A second edge, separately attributed to a curator or research model, may classify the treatment of B as favorable, critical, or procedurally distinguished. An assertion that A overrules B should identify the precise legal issue and the institutional competence required for that claim. It is rarely safe to infer reversal simply from negative language in a paragraph.
Legal citations are directed and contextual. Counting incoming citations may measure attention in a defined corpus, but the numerical pattern alone cannot establish binding force, precedential status, or doctrinal importance. Different legal traditions assign distinct functions to decisions, reports, and judicial reasoning. Corpus incompleteness may amplify apparent centrality, particularly when older or lower-court records are underrepresented.
A researcher answering why a decision was retrieved needs to see the citation trail, the passage, the metadata of both decisions, and the processing step that assigned any treatment label. This is the point at which a graph becomes explanatory rather than a convenient but opaque network visualization.
Normative Authority, Hierarchy, and Jurisdiction
Legal graph relationships operate inside systems of authority. Constitutions, statutes, delegated regulations, administrative decisions, judgments and private instruments derive their normative force from applicable sources and procedures. The hierarchy of those sources depends on jurisdiction, constitutional arrangements, subject matter, institutional competence and time. A cross-national graph that includes all these document categories cannot safely infer one universal ordering of authority.
A relation labeled amends can be documentary and legal when an identified enactment explicitly modifies a prior instrument. A relation labeled overrides demands closer inquiry: does one text prevail because of superior normative rank, later enactment, a special provision, constitutional adjudication, or a jurisdiction-specific conflict rule? Those possibilities have different preconditions and consequences. Recording them under a single edge obscures the basis of the conclusion.
Competence is also conditional. A legislative body's legal power may be limited by territory, subject, delegated authorization, or temporal scope. A court's ruling may bind particular parties, decide a point of law, or carry persuasive authority beyond the immediate litigation. A graph should make the source and procedural posture available to the interpreter rather than transform the graph's connectivity into a hierarchy of courts or enactments.
The graph can represent positive legal authority only through a model that permits source-sensitive reasoning. The ontology supplies categories such as legislative act and judgment; the dataset provides actual texts and judgments; legal analysis specifies the local conditions of binding effect. The OWL 2 overview describes general formal semantics for ontological inferences, but legal hierarchy cannot be read out of ontology classes unless the underlying legal constraints have been modeled and justified.
An illustrative example is a hypothetical municipal rule adopted under a statutory power later narrowed by amendment. Its legal status depends on the empowering law, the amending law's temporal effect, any transitional provisions and the municipal decision's own scope. A query limited to whether the rule was once validly issued will produce a different answer from a query about whether it continues to apply to conduct today.
These distinctions matter especially for cross-border comparison. A shared concept label, identical citation style, or common graph database is evidence of interoperability at one level, not evidence of jurisdictional equivalence. Source-sensitive distinctions should be a first-class part of the query contract.
Multiple Legal Times and Version-Specific Truth
Law has several relevant temporal coordinates. Enactment, publication, commencement, application to events, amendment, repeal, judicial decision, database acquisition, and graph revision can occur on different dates. For a legal proposition to be useful, the researcher must know which temporal dimension the answer addresses. A rule may be enacted in one year, come into force later, and apply differently to earlier transactions under a transitional provision.
The European Publications Office explains in its guidance on consolidated texts that a consolidated text brings an original act together with subsequent amendments and corrections and depicts rules applicable at a point in time. The guidance also explicitly states that consolidated texts have no legal effect: they are documentary aids. This distinction is pivotal for knowledge graphs. A consolidated view can be an excellent navigation interface while the underlying basic act and amendments remain the legal sources to which a claim must be traced.
At least two broad time dimensions arise in a maintained graph. Legal or valid time describes when a rule or relation is applicable in the legal order. Record or transaction time describes when the graph's system knew or stored the assertion. A source discovered in November may describe a legal position that held in June. Backdating the legal interval must not rewrite the audit history of when the source entered the repository. Some datasets require additional time axes for judgments, corrections, source publication, and the period during which an interpretation was accepted.
The researcher should therefore ask temporally complete questions. Instead of “What is the rule?” the query should specify a jurisdiction, subject, source version and legal time. For historical litigation the legally relevant date may be the conduct, filing, decision, statutory commencement, or a combination determined by conflict and transition rules. Temporal schemas should make such choices expressible and prevent an unspecified “current” default from masquerading as an answer to all possible questions.
One useful hypothetical representation associates a provision version with a start date, a recorded end date where established, the legislative event creating or ending the interval, and the source relied on for that interval. An unresolved commencement question should have an explicit unknown or review state, not an invented timestamp. Inclusive and exclusive boundaries must be fixed consistently and documented. Absence of an end date is not necessarily proof of present legal effect.
The result is a basic methodological constraint: any graph answer concerning legal effect through time must be tied to a declared temporal perspective and a traceable source chain.
Amendments, Consolidations, and Legislative Lineage
Legislative amendment generates graph relations that are simultaneously textual and legal. A later instrument may insert a subsection, substitute words, repeal a paragraph, suspend operation, or amend the definition that governs multiple provisions. Amendments can be conditional, delayed, territorial, and subject to savings or transitional arrangements. Consequently, a simple edge from Act B to Act A labeled changes is often too coarse for reliable legal retrieval.
The legal graph should locate a change event at the smallest identified affected text, distinguish the amending instruction from its consolidated effect, and preserve the instruments and dates involved. It should also support the recovery of earlier versions. A version chain is a documented series of representations of the legal text, not proof that each point in the chain was legally operative for every class of circumstances.
The EUR-Lex consolidation explanation identifies amendments, corrigenda and accession-related modifications within consolidated versions. Its explicit warning about the non-legislative status of consolidated views requires a graph to keep legal-source lineage separate from editorial compilation. A consolidated page may display words “as of” a particular date, but the authority for those words comes from the authentic legal acts and applicable legal rules.
A provision may survive in number while its legal content changes several times. A stable provision identifier is convenient for navigation, yet an assertion about the wording effective on a past date should point to a version-specific expression. Conversely, a provision might be renumbered while retaining substantially the same language. A relation describing textual continuity should not be substituted for an analysis of normative continuity.
A useful amendment query asks for the sequence of source instruments and provision-level effects, the operative dates established by those instruments, and any unsettled implementation information. This provides an evidentiary explanation that a simple “last updated” timestamp cannot deliver. It is especially important when automated answers will be used for litigation deadlines, compliance duties, or retrospective legal research.
Provenance, Authenticity, and Auditability
Provenance links a graph assertion to the documentary record and the processes that brought it into the system. It is distinct from an assertion's truth. A reproducible extraction can faithfully reproduce an erroneous source, and an authenticated text can still be misinterpreted. The W3C PROV Ontology provides classes and relationships for entities, activities, agents, derivation and attribution. These tools can document that an assertion was generated from a particular file by a named transformation and subsequently reviewed.
A legal graph's provenance record should ordinarily identify the repository and publisher, retrieval date, canonical source URL, document identifiers, edition or expression, language, relevant passage, acquisition method, source checksum where appropriate, extraction software or human analyst, and correction history. The depth of recording should match the consequence of the assertion. A simple classification for topical navigation may need less procedural detail than an assertion that a penalty provision was in force at the time of an alleged offense.
Authenticity is an institutional and evidentiary question. A file hash proves that specified bytes match a recorded digest. It does not prove that those bytes are the current official text or that the purported issuing authority actually issued them. Source identity should therefore be checked against the publisher's institutional authority and, where applicable, official signatures, publication notices, authentic versions or other legal verification mechanisms. The graph may preserve hashes as part of the audit trail without conflating integrity and authenticity.
Named graphs can be used to group RDF assertions by data source or processing batch; statement-level provenance can use other documented representations. The choice depends on the granularity needed and on compatibility with query and inference software. In all cases the legal research interface should permit inspection of the source of a consequential edge. It should also expose when an assertion originated from automatic extraction and has not received documentary or expert confirmation.
The most revealing audit question is not merely “Who created the node?” but “Which exact source and inferential step support the relationship now presented as law?” A chain of provenance is adequate when it allows an independent researcher to locate the cited source, recover the relevant words and reproduce or contest the transformation that produced the assertion.
Norm Representation and LegalRuleML
A knowledge graph of legal sources and relationships can be used alongside formal representations of norms. Rule representations address predicates, conditions, exceptions, obligations, permissions, temporal qualifications and potentially conflicting rules. The LegalRuleML Core Specification, issued as an OASIS Standard in 2021, extends rule markup to capture features of legal normative reasoning and associated metadata. It occupies a different functional level from a general citation graph.
An RDF edge may say that a statutory section establishes an obligation. A machine-interpretable norm representation seeks to specify who must do what, under which conditions, and with which exceptions. That representation must identify its source, jurisdiction, effective period, and the interpretation embodied in its formalization. It should also record whether the rule is an enactment's text, a court's doctrinal statement, an administrative interpretation, or a model builder's abstraction.
Formalization frequently requires decisions about legal language whose content depends on definitions and context. The phrase “unless otherwise provided” directs attention to possible exceptions elsewhere. “Reasonable” introduces evaluative standards. A statutory “may” might confer a power, permission, or discretion depending on its context. An ontology or rule engine can encode a chosen interpretation, but it cannot make the methodological choice invisible. Each consequential reading must be explicit, inspectable and reviewable.
LegalRuleML supports an interchange standard for certain representations; it does not establish a universal implementation of all legal reasoning. Formal rule execution should be tested against source-grounded fact patterns, exceptions, contradictory provisions and relevant judicial interpretations. Legal graph retrieval can help gather the authorities on which a rule model depends; the rule model can in turn generate explanations whose cited foundation is recoverable through the graph.
The analytical conclusion is that document graphs and norm models are complementary. Their integration has value when the boundary between recorded source relationships and interpreted legal consequences remains visible.
Extraction Pipelines and Generative AI
A graph built from legal text depends on extraction tasks that appear simple only after their outputs are accepted as facts. Typical stages include identifying document boundaries, recognizing named legal entities and citations, resolving references, mapping passages to ontology classes, assigning temporal attributes, and creating relation assertions. Each stage can add uncertainty. A citation parser may detect an expression correctly yet associate it with the wrong authority; a named entity detector may recognize a tribunal but miss that the name refers to its predecessor.
The 2024 legal NLP survey reviews tasks such as named entity recognition, classification, legal question answering and judgment prediction and identifies technical challenges affecting legal text analysis. Its coverage supports analysis of method and limitations; it does not warrant assuming that a specific graph generated for a new corpus is accurate. Extraction should be assessed against manually verified, jurisdictionally appropriate reference data.
Recent work illustrates differing approaches. D'Amato and colleagues (2025) describe two approaches to the construction of a legal knowledge graph, including a bottom-up method and a large-language-model-assisted path, and discuss validation using competency questions. This is evidence of an investigated research program, with its own corpus and reported methods. It supplies neither a universal accuracy guarantee nor proof that unverified generated edges are doctrinally reliable.
Generative systems may propose candidate relations and summarize sources. Their outputs require a trace from the proposed edge to the exact supporting passage and an identifiable verification decision. A model-generated claim that one case overruled another has greater legal significance than a model-generated topical similarity. The review threshold should reflect that difference. If an entity cannot be unambiguously resolved, competing candidate matches should be retained or the edge withheld from authoritative views.
Extraction quality should be measured at several levels: document detection; entity and span recognition; relation classification; identity resolution; temporal association; provenance completeness; and end-to-end legal answer correctness. An aggregate F1 score for a model can conceal catastrophic errors on rare amendment or overruling relations. Error analysis should report false positive relationships that could change the result of legal research, alongside ordinary precision and recall.
Human review is valuable when it tests the connection between a text and a graph assertion rather than merely approving fluent output. A reliable pipeline records revisions to its extraction rules and permits correction of downstream edges affected by a changed source or model. Graph construction is thus an ongoing scholarly and technical process, rather than a one-time conversion of documents into triples.
Validation, Competency Questions, and Quality Assurance
Graph validation begins with explicit statements of what is required from the dataset. The SHACL standard enables an RDF data graph to be checked against shapes defining structural conditions such as cardinalities, datatypes and permitted relationships. An instrument node might be required to have a stable identifier and jurisdiction; a decision could be required to have an issuing court and decision date; a sourced claim could be required to have a source reference and processing provenance.
Passing a SHACL check establishes satisfaction of the specified constraints under the validation procedure. It does not establish that a cited statute exists, that a decision has been characterized correctly, or that a provisional rule has legal effect. Moreover, a constraint can be too weak or itself encode the wrong assumption. Source verification and legal review operate beyond schema validation, even when the graph schema makes their tasks easier.
Competency questions express the practical needs a graph should answer. Examples include: Which instrument versions were applicable to a specified date? Which official act amended a particular provision? What is the provenance of a claim that Decision A cites Decision B? Which cited court decisions have been superseded, and according to what authority? Each question should be accompanied by a declared corpus, jurisdiction, temporal frame, expected output type, and adjudication of ambiguous answers.
A well-designed test suite mixes positive, negative and adversarial examples. Positive tests check known relationships. Negative tests verify that a graph does not infer “overrules” from a mere citation or “in force” from the absence of repeal metadata. Adversarial cases include two courts with identical abbreviations, an amendment enacted but not yet commenced, and a multilingual title that maps to distinct legal concepts. A graph can pass ordinary data-shape tests and still fail these legal-semantic tests.
Tests should distinguish coverage from correctness. Precision measures how often returned relations are correct within an evaluated sample; recall measures how many relevant relations were found. Neither is meaningful without a documented population, source versions, labeling rules, sample design and relevant uncertainty. High query completion speed is not evidence of doctrinal completeness. The need to explain each relation and reproduce its source is especially important when graph output feeds legal advice, adjudication or regulatory work.
The W3C SPARQL specification permits queries against a selected RDF dataset. A defensible legal QA report should record the dataset version and entailment assumptions used by the query. The report should also identify excluded sources and known temporal gaps. Testing is adequate only when it evaluates the actual task the graph purports to support.
Multilingualism and Comparative Legal Semantics
Legal knowledge graphs cross language boundaries through names, citations, official translations and multilingual expressions of instruments. The existence of several language versions of a legal text does not imply that every version has the same legal status. In some legal systems multiple versions may be equally authentic; elsewhere a translation is supplied for information. A graph must preserve the official language and authenticity status of the source expression rather than convert a translated string into a universal identifier of the rule.
European publication infrastructure is especially instructive because it distinguishes works, language expressions and formats. The Publications Office's Cellar metadata model describes multiple language and format versions of publications and their identification. Linking those representations supports cross-language retrieval while preserving the relation between the underlying work and the textual form actually consulted.
Legal terms can also resist one-to-one mapping. “Consideration,” “equity,” “ordre public,” “Rechtskraft,” and other doctrine-specific expressions carry histories and institutional functions that a literal translation cannot settle. A responsible graph may record a translation label for search while attaching an explicit mapping status to any assertion of conceptual similarity. The SKOS mapping vocabulary provides a means of expressing related, broader, narrower and close conceptual correspondences without imposing universal identity.
Comparative research should track provenance for every translation, including whether it is official, professionally supplied, or produced by a machine. A concept node can hold multilingual lexical labels; a legal proposition must remain anchored to a particular source-language text and jurisdiction. Where competing translations matter, preserve both together with notes about the interpretive dispute rather than collapsing them into one fluent paraphrase.
Cross-jurisdictional alignment should be evaluated against actual research questions. A taxonomy suitable for searching employment-law topics across countries may be inadequate for determining the legal consequences of a specific form of contract. Contextual comparison can produce useful links without asserting that the linked concepts are identical. The graph should allow researchers to see the comparative judgment and reconstruct its evidence.
Legal Research Queries and Explainable Retrieval
Graphs are especially valuable when research questions involve chains of identified relations rather than isolated words. A researcher may need all amendments to a provision, all decisions citing a specified instrument, the official language expressions of one legislative work, or the legal authorities associated with a court's reasoning. A source-linked graph can expose these paths and allow the researcher to reopen the original texts.
A sample research query may ask for decisions issued by a particular court that cite an identified provision between two dates. Its correct evaluation depends on reliable identifiers, court metadata, provision granularity, citation extraction, and decision dates. If the graph contains only decisions from one reporting service, the output describes that service's coverage. An explanatory interface should display the restriction as part of the answer, not as a hidden condition of the query.
The European Commission's ECLI identifier system helps supply normalized handles for judicial decisions; ELI performs an analogous role for legislative information. Their coexistence supports cross-document navigation. It does not create substantive legal reasoning automatically. A retrieved path must still differentiate mentions, citations, legal amendments and judicial interpretations.
Explainable retrieval requires the system to identify why an answer was returned. For an exact citation relation, that may be a source paragraph and extracted citation. For a conceptual recommendation, it may be a topical similarity score with stated limitations. For an asserted legal rule, it should include the operative source version and the interpretive steps. A graph should not present a model's semantic neighborhood as though it were a jurisdiction's binding legal relationship.
The data model also matters for negative findings. “No amendment found” may mean none in an official, complete, current register; none in a selected corpus; or none identified by an imperfect extractor. The user must be able to distinguish these claims. A good research query carries its coverage information and produces a response that can be assessed against the underlying source inventory.
European Legislative Linked Data in Practice
The European Legislation Identifier is a real institutional example of linked legal information. The EUR-Lex ELI description identifies three core functions: standardized legislative URIs, metadata, and machine-readable exchange. It also notes that implementation is voluntary across national legislative publication systems. This is an important limitation on any statement that ELI has produced complete European legal interoperability.
ELI's standardized URLs can distinguish legislative act types, years, document numbers and certain document subdivisions when implemented. Their value lies in persistence and shared interpretation. The graph can use these identifiers to relate official legislation, metadata and document versions without manufacturing local identifiers for everything. Yet each linked institution still governs its own source, publication practice and legal status.
The ELI implementation documentation provides publisher-facing guides and describes ontology-based metadata and synchronization mechanisms. These materials support a concrete institutional account of cross-border legislative metadata. The scholarly inference is that graph interoperability requires coordinated identifiers, vocabularies, publication practices and correction processes, with each element providing a distinct component of assurance.
The Publications Office's Cellar supplies a second layer. Its knowledge graph interface stores metadata in a triple store and provides SPARQL queries using the Common Metadata Model. The Cellar data documentation describes retrieval through REST and SPARQL and describes metadata as linked open data. These official accounts establish that an institutional knowledge graph underpins actual publication services.
Cellar's documentation describes work, expression and manifestation relations, which are essential to representing different language and format embodiments of one legal source. That structure illustrates the need to avoid treating a PDF, an expression and a legislative act as the same entity. It also shows how bibliographic models can contribute legal interoperability without attempting to encode every doctrinal relationship.
As a comparative case, ELI plus Cellar demonstrates the value of persistent identifiers and high-quality metadata backed by a publisher. It does not establish that a third-party graph derived from Cellar is complete, free of errors, or legally authoritative. Those properties require independent inspection of data versions, coverage and the particular legal claims asserted by the consuming system.
European Case-Law Identification in Practice
The European Case Law Identifier standard addresses a related but separate coordination problem. The official EUR-Lex ECLI explanation describes the structure of an identifier that combines the ECLI prefix with a country, court, year and distinctive number. The 2019 Council ECLI conclusions set out the related metadata policy. Those records support legal research that spans multiple judicial repositories.
The Court of Justice's case names and citations guidance confirms that the Court uses ECLI for European Union court decisions and explains citation conventions. ECLI can therefore act as a reliable cross-platform reference where correctly assigned by the relevant authorities. Its function is to identify a judgment, not to determine what its holding means.
In a knowledge graph, a decision may have a normalized ECLI alongside national or reporter citations, language expressions, publication URLs and judicial metadata. Citation relations can then be linked to the identified decision even if users enter different textual forms. Accurate resolution may still require court codes and years to distinguish similar short names.
Treatment classification introduces new uncertainty. An appeal may affirm part of a decision and reverse another part; a later judgment may distinguish a proposition while citing the earlier decision approvingly elsewhere. The graph should be capable of associating characterization with a specific passage and issue. A simple positive-versus-negative label is often inadequate for substantive legal interpretation.
European identifier harmonization thus supports the discovery and referencing layer of research. The legal-semantics layer remains dependent on original judgments, procedural setting, national and supranational doctrines of authority, and justified human or machine-supported interpretations.
Legislative Change Tracking in Practice
An official legislative revision service provides a direct test of whether a knowledge graph can distinguish the production of an enactment from the editorial maintenance of its text. The United Kingdom's legislation.gov.uk publisher help documentation describes revised texts, changes that have not yet been applied, prospective provisions, and different meanings of legislative effects. It also documents the practical lag between a legal change and its incorporation into a maintained revised presentation. Those facts expose a methodological hazard: treating the text most recently supplied by a publisher as a complete account of all changes legally in effect.
The UK platform differentiates a provision's revised text from outstanding effects. It describes a prospective version as an amended text for a provision whose amendment has not yet commenced. In a knowledge graph, both the prospective text and the operative text may be useful research objects, but they answer different questions. A query for a hypothetical date on which a new rule is to commence may need the prospective expression; a query for the current applicable provision should account for the commencement position and outstanding effects. The graphical relationship must specify which of these meanings it carries.
The service's Changes to Legislation register records changes including amendments, repeals, modifications and commencement information. A record of a modifying instrument is not necessarily identical with the editorial text resulting from applying that instrument. For graph research the change event should therefore be linked to the originating enactment, the target provision, the type of effect, its date or temporal condition, and the publisher's revision status. A normalized event can be used to navigate to both the amending text and its affected provision.
Consider a purely hypothetical statutory section S. Act A enacts S on January 1 of a given year; Order B brings S into force on April 1; Act C, made on June 1, provides for a substitution to commence on September 1; and a publisher records that substitution in its database on October 15. Five documentary assertions follow: A contains S; B specifies its commencement; C instructs a change to S; the substituted S has an operative interval beginning September 1 subject to the precise terms; and the publisher's record of that event arrived on October 15. The dates are invented to clarify the data structure and make no claim about any actual statute.
Suppose a researcher queries the legal position on August 15 using an October edition of the graph. The answer should return the pre-substitution text, with C recorded as a prospective change. On September 15, the substituted text may be operative even though the publisher had not yet entered its revised presentation. The answer must distinguish the source-established legal position from the repository's state of knowledge at the time. This comparison shows why graph timestamps, editorial revision stamps and legislative commencement times are analytically distinct.
A more difficult hypothetical case arises when the amending instrument includes an exemption for proceedings begun before September 1. The normative effect of the substitution depends not only on a date but also on facts concerning the proceeding. A graph that returns one undifferentiated “in-force section” misses the transition. To answer responsibly, it must retrieve the savings clause or record a temporal qualification whose meaning is supported by the source.
The practical research implication is a discipline of reconstruction. Automated change tracking can identify the sources and events that require attention, but the answer to an applicability question must integrate commencement, transition, territorial extent, and applicable interpretive law. The graph should expose the path to those materials and indicate when its source coverage is incomplete.
Privacy, Access Control, and Institutional Responsibility
Legal texts often concern identifiable individuals, private organizations, sensitive allegations, criminal proceedings, children, confidential disputes or protected litigation records. Linking dispersed public records can increase the ability to identify people even if each source was individually accessible. Graph technologies are especially effective at creating relational information that was difficult to obtain from isolated records. Institutional responsibility for the graph therefore extends beyond the legality of downloading its source documents.
In the European Union, Regulation (EU) 2016/679 requires, subject to its applicability and the relevant legal bases, lawful processing, purpose limitation, data minimization, accuracy and accountability. The official General Data Protection Regulation sets out these principles in Article 5 and the legal bases in Article 6. A knowledge-graph operator processing personal data needs to determine which law applies to its actual operations, which source restrictions govern it, what the purpose of the linkage is, and whether disclosure of the combined dataset is justified. A citation to GDPR alone cannot establish the compliance status of every cross-border legal research system.
Public legal records can also have publication restrictions or require anonymization. A graph of decisions should preserve authoritative restrictions as metadata and ensure that derived nodes and edges do not inadvertently reveal identities removed from the published text. Anonymization at the document level may be undermined by combining rare attributes and relational paths. A publication system can legitimately withhold a relation or aggregate results even when the underlying record is held for lawful research.
The graph's access-control model may need to distinguish public identifiers and metadata from restricted text, annotations, internal review notes and personal data. Query permissions should apply to inferred relationships as well as to raw documents: concealing a source document while revealing its precise participant relationships may defeat the restriction. Logging access to sensitive records, assessing inference risk and establishing retention rules form part of data governance.
Accuracy is an ethical and legal issue in this domain. An incorrectly merged entity may associate a person's name with an unrelated proceeding. In systems affected by data-protection law, correction procedures should include the ability to trace and repair downstream relationships. A meaningful audit log records the prior assertion, its provenance, its correction, and the effect on published or cached graph outputs.
The appropriate substantive conclusion is that legal graphs are institutional information systems. Their design involves copyright and database licensing, data protection, rights of access, and the public function of legal information. The rules are jurisdiction- and context-specific; a universal publication setting cannot be derived from the technical openness of RDF or the public accessibility of one source.
Governance, Persistence, and Knowledge Maintenance
A knowledge graph remains useful only if its identity, coverage and correction rules survive changes in documents, institutions and software. Publication services must reconcile archived source versions with new official data, migrations of court and legislative identifiers, metadata corrections, newly discovered amendments and modifications to legal ontologies. A stable namespace is an editorial and technical commitment that requires ongoing maintenance.
The DCAT 3 Recommendation specifies interoperable descriptions of datasets and data services and includes terms for versioning, dataset series and checksums. These tools can help register what a legal graph contains, where its distributions reside, and which versions were examined in a research study. A catalogue's metadata does not substitute for the source's legal authority, but it supplies a necessary description of the corpus on which conclusions depend.
Maintenance decisions should distinguish a correction to an erroneous graph assertion from a change in the law itself. If a researcher discovers that a prior citation was incorrectly resolved, the correction belongs in the graph's editorial history. If the legislature amends the cited statute, the change belongs to the law's substantive history, with a corresponding update to the graph. A third kind of change arises when a publisher replaces a PDF or changes a URL without changing the substantive law. These events require separate audit records.
Versioning also concerns ontology and query definitions. If a later vocabulary version makes a concept narrower, previously saved query results may change even if the same underlying legal sources are present. Reproducible research must retain the ontology version, data version, inference settings and query text. A method that cannot reconstruct those dependencies can explain the current graph but cannot securely reconstruct earlier research results.
The governance role of an official publisher differs from the role of an academic research team. A publisher is institutionally responsible for the records it issues and the metadata it maintains under its governing framework. A research graph may combine the publisher's authoritative facts with scholarly classifications and automated inferences. The latter should remain attributed to their creators. Institutional trust is not transitive: an official source linked to a generated proposition does not make the generated proposition official.
A sustainable editorial policy defines procedures for source ingestion, authoritative-status checks, corrections, appeals against classification, access restrictions, and reproducible releases. It specifies how an unsupported assertion can be retired while preserving the history of research that cited it. This allows the graph to be revised without making earlier scholarly use impossible to audit.
Evaluation, Error Analysis, and Research Reproducibility
A graph's performance should be evaluated against the scholarly tasks it claims to support. A project devoted to amendment research should test whether changes, effective dates, provisions, and transitional effects are recovered correctly. A project for judicial citation search should test whether each reference resolves to the appropriate decision. A project for cross-lingual conceptual navigation should evaluate whether its mappings preserve meaning in the relevant legal systems. Results on one task are not a demonstration of performance on another.
A defensible evaluation corpus must be described by source, territory, language, date range, document type, sampling method, and exclusions. The benchmark should preserve source passages and the rules by which reviewers labeled relationships. If reviewers disagree about an interpretation, the disagreement is data about the task. It should not be silently resolved by treating one label as self-evident ground truth.
Data quality and legal quality require different measurements. For entity resolution, precision and recall against verified references are appropriate. For a legal applicability answer, the assessment also needs the correct jurisdiction, operative dates, exceptions, source authority and treatment of uncertainty. For provenance, it must establish whether a reviewer can recover the original text and transformations that produced a contested edge. For reproducibility, another researcher should be able to run the archived query against the same identified graph and inference configuration.
A practical test matrix should include cases that deliberately defeat simplistic assumptions. A provision that has been renumbered tests identifier continuity. Two unrelated cases with similar short citations test entity disambiguation. An enacted but uncommenced amendment tests temporal reasoning. A consolidated non-authentic document tests separation of source authority from readability. A translated term that is only partly equivalent tests conceptual mapping. A decision citing but rejecting another judgment tests the difference between a citation edge and doctrinal treatment. Each is a representative test class, not a claim of a measured result from this study.
Coverage deserves a place alongside accuracy. A graph might have high relation precision because it contains only easily parsed newer cases while excluding historical decisions and unreported judgments. A user should know the composition of the graph and the categories it omits. The decision to return “unknown” can be more accurate and legally responsible than completing a missing relationship from semantic similarity alone.
Where learned extractors are used, model performance should be assessed with held-out legal sources and independently labeled examples, while documenting changes in statute, case reporting and terminology that may affect performance over time. For a graph built with a language model, source grounding and hallucination analysis are necessary, and exact citations should be rechecked. The legal language-processing survey supplies the methodological context for recognizing task-specific evaluation and data constraints.
This monograph does not report an empirical experiment or claim measured gains over legal search engines. Its conclusions concern the conditions under which future empirical claims would become methodologically credible. The separation matters because a technically coherent model, an established interoperability standard, and a tested legal information service are different kinds of evidence.
Comparative Findings and Analytical Synthesis
The preceding comparison yields a set of analytically connected conclusions about legal knowledge representation. The first concerns identity. ELI, ECLI, Akoma Ntoso and Cellar address overlapping but distinguishable objects: legislation and metadata, decisions and citations, structured documents, and multilingual publication representations. Legal knowledge graphs are strongest when they assemble these identifiers through explicitly typed relations while preserving the distinct source models from which they originate. Collapsing their referents into generic document nodes removes precisely the distinctions needed for legal research.
A second conclusion concerns evidentiary depth. Some graph assertions can be checked directly by inspecting a source document: a section belongs to an identified act, a judgment contains a citation, or an institutional publisher supplies a named language version. Other assertions are doctrinal interpretations: a case controls a question, a provision applies to a class of past events, or a legal concept in one system is equivalent to a concept in another. The graph should support both kinds of research, but the evidence and level of review appropriate to them differ.
A third conclusion concerns time. The legal system changes through acts and decisions, while the research infrastructure changes through acquisition, revision and correction. The United Kingdom's official distinction between revised provisions and outstanding effects, and EUR-Lex's distinction between authentic acts and consolidated documentary aids, show that publication state is not identical with legal effect. Each has to be represented with its own source and temporal meaning. A query is legally informative when its answer declares the temporal perspective it implements.
A fourth conclusion concerns authority. A graph relation may connect authoritative legal texts without itself being an authoritative statement of law. The institutional quality of a source cannot automatically transfer to an automated classification, generated interpretation or machine inference. Source attribution, expert review and the explicit recording of contested classifications are the practical means of preserving that distinction. This is especially important when outputs are presented through natural-language interfaces that hide underlying relation types.
A fifth conclusion concerns verification. RDF, OWL, SPARQL and SHACL provide formal tools for representation, inference, query and constraints. These techniques make certain classes of data error detectable and certain paths reproducible. They leave substantial legal questions to source analysis, doctrinal interpretation and empirical testing. A complete QA process must consequently connect schema conformance to source authenticity, provenance, semantic correctness, temporal validity and actual task performance.
The synthesis also identifies a general relationship between interoperability and institutional responsibility. Linked data makes separate legal corpora searchable together, but responsibility for each source and for each derived assertion remains distributed. The appropriate scholarly posture is to preserve those boundaries in the graph and in the user-facing explanation. Interoperability increases the value of a traceable source chain; it also raises the cost of an untraceable mistake because a single false merge can propagate across many connected queries.
These are conclusions drawn from comparison of the documented standards and institutional practices, rather than a proposal for a new named theory, ontology or normative doctrine. Their significance is practical and methodological. They show how the quality of a legal graph can be judged by the kinds of propositions it enables a researcher to verify, reconstruct and contest.
Research Boundaries and Further Work
The monograph analyzes public documentation of established standards and selected legal publishing infrastructures available at the stated research cutoff. It does not contain a newly implemented cross-jurisdictional graph, an experimental dataset, measured extraction performance, full-scale legal annotation study or independent legal opinion concerning a live dispute. It therefore makes no claim that a particular technical pipeline has already achieved the reliability criteria described here.
The strongest documentary case studies concern European and UK legal information infrastructures, with W3C and OASIS standards providing the technical foundation. This selection supports examination of identity, metadata, versioning and provenance under well-documented publication practices. It does not support universal assumptions about jurisdictions with different judicial reporting practices, nonpublic source registers, oral legal traditions or distinct evidentiary regimes.
Access limitations also affect the literature comparison. Official specifications, guidance pages and accessible article metadata provide evidence for particular standards and research positions. Where an academic full text was not examined, the treatment is limited to a verified abstract, bibliographic record or otherwise observed research description. Claims of achieved system performance and legal accuracy are withheld where corresponding data were not reviewed. The source ledger records these distinctions for the present edition.
A future evaluation could construct a reproducible benchmark around official amendments, court citations and temporal queries in two specified legal systems. It could compare reference resolution, passage-level provenance, graph completeness and error severity under human and machine-assisted extraction. Such a study would require licensing review, a documented corpus, gold-standard labeling, independently reviewed legal interpretations and a clear deployment context before reporting quantitative conclusions.
The key maintenance questions are equally concrete: how an institution should retire erroneous relationships without erasing audit history; how to propagate a corrected source to derived graph assertions; and how to preserve query reproducibility as ontologies, legal data and model versions change. These are research problems whose answers require evidence from operational systems. They are identified here as future work, not offered as findings from experiments that have not occurred.
Conclusion
Legal knowledge graphs organize law into explicit and traversable assertions about documents, institutions, concepts, sources and change. Their most defensible contribution to legal research is the ability to identify relationships and show how they were established. Persistent identifiers, structured legal texts, interoperability vocabularies, query languages and provenance records provide powerful means of making dispersed legal information discoverable and inspectable.
The comparative inquiry establishes that graph representation must preserve the identity and version of each legal source, separate documentary relationships from interpretive classifications, register multiple legal times, retain jurisdiction and source authority, and expose how assertions were created and corrected. Legislative revisions, judicial citations and multilingual concept mappings each illustrate why these distinctions determine the reliability of a research answer.
The source and graph layers remain mutually connected. An official act or judgment establishes the documentary foundation; a graph encodes propositions about it; a scholarly interpretation evaluates the legal meaning; and validation tests specified characteristics of the resulting representation. A responsible legal research system permits the researcher to return from an asserted relationship to the source, date, passage and reasoning on which it depends.
A legal knowledge graph succeeds, in scholarly terms, when it strengthens the verifiability, contextual precision and reproducibility of research across sources and over time. That standard accommodates advancing computational methods while maintaining the discipline of legal interpretation. It provides a clear basis for assessing future legal graphs through their recorded sources, transparent methods, documented limitations and demonstrated accuracy.
References
- Ariai, Farid, and Gianluca Demartini. 2024. Natural Language Processing for the Legal Domain: A Survey of Tasks, Datasets, Models, and Challenges. arXiv:2410.21306. Source
- Ashley, Kevin D. 2017. Artificial Intelligence and Legal Analytics: New Tools for Law Practice in the Digital Age. Cambridge: Cambridge University Press. Source
- Council of the European Union. 2019. Council Conclusions on the European Case Law Identifier (ECLI) and a Minimum Set of Uniform Metadata for Case Law (2019/C 360/01). Official Journal of the European Union C 360, October 24, 1–22. Source
- Court of Justice of the European Union. n.d. Case Names and Citations. CURIA. Source
- D'Amato, Claudia, Giuseppe Rubini, Francesco Didio, Donato Francioso, Fatima Zahra Amara, and Nicola Fanizzi. 2025. Automated Creation of the Legal Knowledge Graph Addressing Legislation on Violence Against Women: Resource, Methodology and Lessons Learned. arXiv:2508.06368. Source
- European e-Justice Portal. n.d. European Case Law Identifier (ECLI). Source
- European Parliament and Council. 2016. Regulation (EU) 2016/679 of 27 April 2016 (General Data Protection Regulation). Official Journal of the European Union L 119, May 4, 1–88. Source
- Hoekstra, Rinke, Joost Breuker, Marcello Di Bello, and Alexander Boer. 2007. The LKIF Core Ontology of Basic Legal Concepts. In Proceedings of the Workshop on Legal Ontologies and Artificial Intelligence Techniques (LOAIT 2007), 43–64. Source
- OASIS LegalDocML Technical Committee. 2018. Akoma Ntoso Version 1.0. Part 1: XML Vocabulary. OASIS Standard. Source
- Palmirani, Monica, Guido Governatori, Tara Athan, Harold Boley, Adrian Paschke, and Adam Wyner, eds. 2021. LegalRuleML Core Specification Version 1.0. OASIS Standard, August 30. Source
- Publications Office of the European Union. n.d. Cellar Data: Knowledge Graph. Source
- Publications Office of the European Union. n.d. Cellar Data: Metadata. Source
- Publications Office of the European Union. n.d. Cellar Data. Source
- Publications Office of the European Union. n.d. Consolidated Texts. EUR-Lex. Source
- Publications Office of the European Union. n.d. European Case Law Identifier (ECLI). EUR-Lex Help. Source
- Publications Office of the European Union. n.d. European Legislation Identifier (ELI). EUR-Lex Help. Source
- Publications Office of the European Union. n.d. Implementing ELI. EUR-Lex ELI Register. Source
- The National Archives (United Kingdom). 2026. Help: Revised Legislation, Prospective Versions, and Outstanding Changes. legislation.gov.uk (information current at November 21, 2026). Source
- The National Archives (United Kingdom). n.d. Changes to Legislation. legislation.gov.uk. Source
- World Wide Web Consortium (W3C). 2009. SKOS Simple Knowledge Organization System Reference. W3C Recommendation, August 18. Source
- World Wide Web Consortium (W3C). 2012. OWL 2 Web Ontology Language Document Overview (Second Edition). W3C Recommendation, December 11. Source
- World Wide Web Consortium (W3C). 2013. PROV-O: The PROV Ontology. W3C Recommendation, April 30. Source
- World Wide Web Consortium (W3C). 2013. SPARQL 1.1 Query Language. W3C Recommendation, March 21. Source
- World Wide Web Consortium (W3C). 2014. RDF 1.1 Concepts and Abstract Syntax. W3C Recommendation, February 25. Source
- World Wide Web Consortium (W3C). 2017. Shapes Constraint Language (SHACL). W3C Recommendation, July 20. Source
- World Wide Web Consortium (W3C). 2020. JSON-LD 1.1: A JSON-Based Serialization for Linked Data. W3C Recommendation, July 16. Source
- World Wide Web Consortium (W3C). 2024. Data Catalog Vocabulary (DCAT), Version 3. W3C Recommendation, August 22. Source