Law by BogdanovaThe Encyclopedia of Law and Legal Studies

RESEARCH DOSSIER

Legal Corpus Linguistics

Abstract. Legal corpus linguistics uses systematically assembled evidence of language use to investigate questions arising in the interpretation of statutes, constitutions, judgments, regulations, contracts, and other legal texts. Its evidentiary contribution depends on whether the sampled language reflects the relevant communicative community, genre, jurisdiction, historical period, and legal question. This monograph examines the theoretical premises, linguistic instruments, corpus construction, inferential procedures, judicial uptake, methodological disputes, diachronic and multilingual applications, and conditions of reproducibility. It closely analyzes Muscarello, Taniguchi, Rasabout, Boler, and Garibay while distinguishing the holdings of courts from separate opinions and scholarly reconstructions. The literature presents both a significant opportunity to test judicial intuitions and a warning that observed usage is not equivalent to the full extension of a legal concept. The central finding is methodological: a corpus analysis establishes a qualified proposition about a specified distribution of attested uses; the transition from that proposition to a legal conclusion requires an independently articulated interpretive rule and a check against the operative legal sources. The study provides an auditable research procedure, detailed examples of failure modes, and a comparative treatment of multilingual legal language. The contribution is a synthesis and evaluation of existing methods, rather than a claim to have executed an original linguistic experiment or to have created a new interpretive doctrine.

Keywords: legal corpus linguistics; statutory interpretation; ordinary meaning; original public meaning; collocation; concordance; corpus representativeness; historical semantics; judicial reasoning; multilingual law; reproducible legal research; interpretive evidence

Introduction

Legal interpretation repeatedly asks what language communicates when it is deployed in an authoritative legal setting. A word can have several conventional senses, and a phrase can be recognizable in conversation while carrying a specialized sense in a statute. Legal corpus linguistics investigates these questions through collections of attested speech or writing, frequently searchable by computer. Rather than selecting a memorable example of a word, the researcher identifies the linguistic population relevant to the dispute, examines an organized sample of occurrences, and makes explicit the steps by which those observations bear on a proposed reading.

The methodological attraction is considerable. Human linguistic memory is selective; legal dictionaries abstract away from particular contexts; judicial opinions often invoke what a word ordinarily means without specifying the empirical basis for that impression. Lee and Mouritsen (2018) argue in Judging Ordinary Meaning that corpus methods can make some such inquiries more systematic and contestable. Their proposal belongs to a larger movement toward evidence-based inquiry into language, including the effort described in Mouritsen's account of an evolving interpretive framework. These works supply a central intellectual case for the method, although a persuasive case for measuring usage leaves open the normative question of which linguistic evidence the law should privilege.

The research question here is precise: under what conditions can distributions of usage support a legally relevant claim about contested meaning, and how should a decision-maker evaluate competing corpus designs? The inquiry concerns both descriptive reliability and juridical authority. A body of texts may be an excellent linguistic dataset while being a poor representation of the persons to whom a statute addresses its rule. Conversely, an official statute archive may be authoritative as a record of law while consisting almost entirely of a highly specialized register. Matching source to question is therefore the foundation of validity.

The present research is a documentary and comparative methodological analysis. Its primary legal evidence consists of published judicial opinions and official institutional descriptions of legal records; its secondary evidence consists of scholarly linguistic, experimental, and jurisprudential work. Direct corpus querying, independent annotation of concordance lines, participant testing, or new prevalence estimates are outside this study's performed methods. Numbers reported from a judicial opinion or institutional corpus description are identified as those sources' reports. Examples created for exposition are marked illustrative and should never be read as observations from a downloaded dataset.

The United States provides the richest judicial case studies in this dossier because the legal corpus debate has appeared explicitly in published American opinions. European Union materials supply a contrasting multilingual setting with a different set of interpretive requirements. The doctrinal proposition drawn from either setting is kept within its jurisdiction and applicable date. An inference about the meaning of an American state statute does not govern interpretation of an EU regulation, and usage in modern American journalism cannot automatically establish eighteenth-century constitutional vocabulary.

The Object and Status of Corpus Evidence

A corpus is a deliberately collected body of language data. Depending on its design, it may contain digitized printed texts, transcripts of speech, translated parallel documents, judgments, legislative enactments, or mixed genres. A corpus is a research instrument whose usefulness depends on provenance, sampling, metadata, annotation, and search behavior. Its size matters only in relation to those properties. A collection with millions of poorly dated words can be less informative for a historical enactment than a smaller dataset whose documents can be dated, attributed, and classified reliably.

Corpus evidence is normally observational. It records that an expression occurred in a source under some identified circumstances. The observation may be stated as a token count, a rate per million words, a list of concordance contexts, a temporal distribution, or a manually classified set of senses. Each is a different type of proposition. None is identical to a legal rule. The legal rule comes from the authoritative text and the interpretive practices recognized by the competent legal system. A linguistic measurement can inform the understanding of those texts; it cannot itself amend them.

This difference matters for documentary source criticism. The U.S. Government Publishing Office describes the Statutes at Large as a permanent session-by-session collection of enacted federal laws and resolutions. It is relevant to determining what Congress enacted and when, as well as to collecting language from statutes. Yet a study of how ordinary speakers used a disputed expression might require newspapers, conversation, fiction, or letters beyond the Statutes at Large. The same source can be foundational for one question and inadequate for another.

It is helpful to distinguish four objects of inquiry. First is the textual object: a particular phrase in an identified and dated instrument. Second is the linguistic population: the speakers, writers, or genres whose usage is being studied. Third is the observational corpus: the actually obtainable sample. Fourth is the juridical question: the legal issue for which the linguistic observation is offered. A valid investigation states all four and explains each connection. Confusing the second and third objects is a sampling error; confusing the third and fourth is a legal inference error.

Theil's (2025) analysis of doctrinal and empirical methods is especially useful here. Legal doctrinal research must attend to sources capable of determining the content of legal rules; empirical work can clarify the meaning, operation, or consequences of those sources while remaining subject to constraints of legal relevance. This study applies that distinction throughout, without assuming that the same rules of legal authority obtain in every jurisdiction.

Intellectual History and Interdisciplinary Foundations

The appeal to actual language use long predates digital corpora. Lexicographers historically assembled citations of words in context; lawyers examined contemporary examples, literary sources, dictionaries, and legal precedent. Computing changed the scale, retrieval speed, and repeatability of the enterprise. It also introduced new problems: electronic collections can aggregate genres whose meaning practices differ, and search interfaces may hide decisions about tokenization, filtering, or variant spellings.

Mouritsen (2017) framed corpus linguistics in legal interpretation as an emerging practice that could make ordinary-meaning reasoning more transparent while raising difficult theoretical questions. Lee and Mouritsen (2018) then developed a more systematic account of ordinary communicative content and applied it to familiar statutory problems. In response to objections, Lee and Mouritsen's (2021) The Corpus and the Critics defended corpus evidence while refining questions of community, context, and method. That exchange shows a developing research practice responsive to criticism, rather than a settled mechanical test that every court accepts.

The intellectual setting extends beyond United States textualism. Corpus linguistics intersects with lexical semantics, sociolinguistics, historical linguistics, quantitative discourse analysis, translation studies, and legal informatics. Wright's (2026) survey of corpus perspectives on legal discourse situates recent work across several legal traditions, emphasizing comparative design, genre networks, and the relationship between texts and institutions. This broader field studies how legal language is produced and circulated, not exclusively how judges decide which sense of a disputed noun should prevail.

A substantial doctrinal distinction nevertheless persists. In statutory construction, a court may ask which meaning an enacted phrase had for a specified community at the enactment date. In forensic linguistics, an expert may examine authorship, threat language, or speech evidence. In institutional discourse analysis, a scholar may study how judges express authority or how legislation names social actors. These inquiries can employ the same concordance tool, but they have different dependent variables and different grounds of validation. Treating their outputs as interchangeable obscures both their strengths and their limits.

Corpus Selection and Representativeness

A defensible corpus begins with an explicit target population. The investigator specifies period, region, language variety, document genre, audience, and practical reason why those attributes correspond to the disputed legal text. A corpus collected for modern television subtitles can help describe contemporary colloquial patterns. It cannot straightforwardly establish meanings in a historical legislative debate. Broad genre balance can reduce one source of distortion, but it does not make a sample representative of every conceivable legal audience.

The Corpus of Contemporary American English documentation describes a roughly billion-word collection drawn from several genres across 1990–2019. The Corpus of Historical American English description identifies historical coverage reaching from the nineteenth century into the 2010s, with changes in source forms and digitization methods across periods. These large and valuable resources remain bounded by their construction decisions. An enactment from 1789 requires sources contemporaneous with that period, not an automatic extrapolation from an 1820-onward historical sample.

In assessing representativeness, ask whether an apparent change in usage is a change in speakers' language or a change in the mix of documents the archive contains. Historical collections regularly combine surviving newspapers, novels, elite correspondence, digitized books, and official papers. The source mixture is shaped by archiving, survival, copyright, and optical character recognition. A word more frequent in a decade with more fiction may reflect genre composition rather than semantic change. A direct comparison must therefore control, where practicable, for genre and source distributions.

The presence of a document in a dataset is not equivalent to independent observations by different speakers. A widely reprinted opinion, syndicated article, or parliamentary statement may appear repeatedly. Counting each copy as a separate event can inflate a linguistic pattern and its apparent certainty. The analysis needs document IDs, deduplication rules, and an account of reprints. Where the unit of statistical inference is a document or author, token-level counts cannot be treated as statistically independent observations.

The University of Chicago Law Review essay The Corpus and the Courts by Tobia (2021) scrutinizes corpus choices in judicial practice and illustrates the hazards of treating historical collections as representative without inspecting their contents. The broader lesson is to publish a collection profile and to show how results change when prominent sources or genres are omitted. Representation is an argument supported by metadata, not a self-validating label such as balanced or large.

Historical Language, Versions, and Temporal Alignment

The relevant date must be justified by the applicable interpretive theory and legal instrument. For a historical ordinary-meaning inquiry, enactment-period sources will frequently carry special weight. In constitutional interpretation, an original-public-meaning question may designate a ratification period and a historical community. In a statute repeatedly amended, the specific wording, enactment history, and any incorporation or reenactment rule matter. A current codification is not automatically identical to the language and arrangement at every past date.

Historical research must also recognize spelling variation and morphological change. A literal string search may miss historical alternatives or inflected constructions. Modern stemming can accidentally conflate words whose senses differed in earlier periods. Researchers should maintain both an unnormalized transcription and a searchable derivative, record corrections, test variant queries, and preserve the date and edition of each source. Automated normalization is a convenience, not a substitute for historical language expertise.

The GovInfo Statutes at Large collection provides an official reference for federal session laws, but the versions found in a legal search engine may reflect later coding, editorial headnotes, or an amended database record. A corpus intended to reconstruct past meaning needs a reproducible link from its tokenized sample back to the particular document version. This is especially important where the legal dispute turns on one historically enacted phrase and subsequent law uses an apparently similar phrase differently.

Archive silence is an additional limitation. A word's absence from a sample may result from lost records, unsearchable scans, genre exclusions, or a genuinely rare communicative situation. Inferences from absence require far stronger assumptions than observing a clear attested use. The more precise claim is often that the researcher found no instance within a stated query and corpus version, after listed filters. That is a reproducible statement even when a broader claim of semantic impossibility is indefensible.

Concordances, Collocations, and Semantic Classification

A concordance displays a target expression and its surrounding context. It is typically the basic unit of close inspection. A keyword-in-context line can reveal that the same word refers to physical movement, possession, administrative action, or an idiomatic expression. Yet a narrow concordance window can conceal a sentence-level negation or a paragraph-level legal definition. A robust protocol therefore permits expansion to the full document and records each annotation decision.

Collocation analysis asks which words occur unusually often near a target term relative to some expected frequency. For legal interpretation, lexical associations may reveal a specialized phrase or dominant usage environment. A collocate such as negligent near misconduct can be relevant, but its meaning depends on whether the term is being contrasted, negated, modified, or defined. An automated ranking gives candidates for interpretation; their classification depends on actual contexts and an explicit research question.

Searches should retain phrase boundaries, spelling variants, inflectional forms, case sensitivity, and alternatives involving conjunctions or prepositions. Gales and Solan's (2020) study of a minister and a laborer reexamines Holy Trinity Church v. United States through historical and legal corpora and reports that variations in the phrase labor or service change the apparent range of use. Their study illustrates why the literal wording of a legal construction, not an isolated keyword, should organize the inquiry.

A semantic annotation scheme needs written categories. For example, in a hypothetical analysis of charge, an annotator may distinguish an accusation, a financial fee, an electrical quantity, and an instruction to act. Each category should be defined with examples and borderline rules. If the legal issue concerns the difference between two senses, the adjudicator should not invent categories after inspecting which classification favors a desired outcome. Ambiguous examples should be retained as ambiguous, rather than forced into the preferred bucket.

When several researchers classify lines, agreement should be measured and disagreements reconciled. The study should record the proportion of indeterminate examples; a result that obtains a majority only by excluding difficult cases is sensitive to the exclusion rule. Automated word-sense disambiguation may assist triage, but its predictions require human error analysis on the relevant genre and historical period. Software accuracy on modern news is not a guarantee of correctness on eighteenth-century legislation.

Frequency, Probability, and the Limits of Inference

A raw number of hits answers a narrow question: how many occurrences a search interface returned under certain conditions. It does not immediately answer how likely a usage was within the target population. Document length, period size, different spellings, duplicate texts, unavailable files, and interface ceilings can affect the count. Rates per million words help compare corpus slices with different sizes, but they do not solve genre imbalance or the normative meaning problem.

Suppose an expressly fictional sample of 100 independently reviewed usages contains 70 examples classified as sense A, 20 as sense B, and 10 indeterminate cases. A researcher can accurately describe the observed coding. If the original pool was deliberately enriched for documents expected to display sense A, however, the 70% observed share is not an unbiased population prevalence estimate. If the 100 examples came from only two authors, a conventional standard error based on 100 independent speakers would be misleading. If sense B is legally admissible even when rare, its lower frequency may be irrelevant to a provision's extension.

The distinction between a random sample and a convenience sample is therefore crucial. A genuine random design describes the sampling frame, randomization procedure, and exclusion rules. A manually selected collection of persuasive examples should be labeled illustrative. Statistical uncertainty is only one dimension of uncertainty: classification disagreement, corpus mismatch, textual corruption, and doctrinal ambiguity may outweigh sampling error.

The most defensible statistical reports specify denominators and comparison baselines. For collocations, a measure of association can be strongly affected by rare words. For time trends, changes in digitization or source genre must be tested. For semantic percentages, uncertainty should include a range of plausible decisions concerning ambiguous items. Any apparent precision should correspond to an auditable method rather than the visual authority of a decimal place.

A central inference error is the transposition of a linguistic likelihood into a legal entitlement: most speakers used the word this way, therefore the court must adopt that reading. The second proposition follows only if the jurisdiction's interpretive practice grants the first the necessary legal significance and if the relevant evidence is representative. Corpus research is most valuable when those connecting assumptions are disclosed.

Research Design and Reproducibility

An auditable study begins with a preregistered or at least dated research question and a written search protocol. The protocol identifies the exact legal language, the source instrument and version, proposed senses, relevant community, corpus platforms, date filters, search terms, units of observation, treatment of duplicates, sample selection, annotation guidelines, and rules for classifying uncertain results. When judicial litigation has already begun, disclosure of departures from the initial protocol is particularly important.

A reproducibility package should include the corpus name and version, retrieval date, stable source IDs, query syntax, output limits, anonymization decisions, random seed where sampling is used, code or spreadsheet formulas, and a complete account of manual exclusions. Licensing may prevent distribution of full source texts. In that event, publish lawful excerpts or source locators, analysis code, metadata, and sufficiently detailed retrieval instructions. A study should state when an online corpus updates dynamically and exact replication cannot be promised after a later database revision.

Quality assurance works at several levels. The technical check tests that a query ran and returned the specified rows. The linguistic check tests whether those rows actually instantiate the target construction and whether their senses are classified consistently. The historical check tests dates and editions. The legal check tests whether the comparison corpus and question fit the operative legal doctrine. These checks should be recorded separately; success in one category is not evidence of success in another.

An adversarially valuable report allows another party to rerun a different defensible corpus selection. Sensitivity analysis should vary the date window, genres, phrase boundaries, duplicate handling, and treatment of ambiguous contexts. If conclusions reverse under small justified changes, the instability is a substantive finding. A legal decision-maker should be able to see that instability rather than receiving a single favorable frequency as an allegedly conclusive result.

Ordinary Meaning and Experimental Validation

The strongest challenge to naive corpus inference comes from direct comparison with human judgments. Tobia's (2020) experimental assessment of ordinary meaning tested the relationship among dictionary definitions, corpus-based examples, and judgments by participants including lay respondents, law students, and judges. The reported divergence undercuts the assumption that a corpus count directly measures what a representative reader believes a word includes. His results do not demonstrate that corpus searches have no value; they demand a more exact account of what the search actually estimates.

A response to this challenge appears in Lee and Mouritsen's (2021) reply to the critics. They maintain that systematic observation of language supplies evidence that can be challenged, refined, and interpreted with context. The dispute turns in part on the target: is the legal inquiry concerned with the most frequent conventional use, a reasonable interpretation in context, what a layperson judges possible, the communicative content likely conveyed at enactment, or the semantic extension selected by law? Different empirical instruments are capable of measuring different objects.

Experimental work also presents its own design choices. An unaided participant's response to a stylized hypothetical may be sensitive to instructions, framing, and whether the participant considers the social norm or legal rule in the background. A corpus may be sensitive to published genre, syntax, and frequency. Neither method escapes construct validity. The researcher must demonstrate that the target in the experiment matches the target in the legal question.

The methodological conclusion is that triangulation can improve evidence. A corpus may illuminate attested contexts; dictionaries may map conventional sense inventories; experimental judgments may test category boundaries; doctrinal analysis may clarify what counts as legally relevant. Their disagreements should be explained as potential evidence of distinct meaning dimensions, rather than discarded as noise. A single source of language evidence rarely settles all those dimensions at once.

Early United States Judicial Encounters

The Supreme Court's decision in Muscarello v. United States, 524 U.S. 125 (1998), considered whether the statutory phrase carries a firearm reached a firearm conveyed in a vehicle. The majority read the phrase broadly, using dictionaries and examples of ordinary usage; the dissent advanced a narrower construction. This was not a controlled contemporary corpus study of the kind later scholars proposed. It remains important because both the interpretive question and the reliance on examples of usage exposed the need to explain how such examples are selected and compared.

The Supreme Court later decided Taniguchi v. Kan Pacific Saipan, 566 U.S. 560 (2012), construing compensation of interpreters in 28 U.S.C. § 1920(6). The majority concluded that the relevant category concerned oral interpretation rather than written document translation, reasoning from contemporary dictionaries and statutory context. The dissent identified usages and practical arguments supporting a broader understanding. The case is frequently used in corpus scholarship as a test of what more systematic evidence might add. The holding must not be mistaken for a court's adoption of the later scholars' reanalysis.

Lee and Mouritsen (2018) use these statutory interpretation problems to show why a dictionary's inclusion of a definition is not the same as information about a word's likelihood in a specified phrase and setting. Their comparison also illustrates how an analyst can reach a different semantic emphasis from a conventional judicial opinion while accepting the underlying facts and procedural posture. Scholarly reconstruction is an argument to evaluate, not a retroactive amendment of the court's legal reasoning.

The cases clarify a recurring issue: a court can ask both whether a reading is linguistically conceivable and whether that reading is ordinary in the relevant statutory context. If those propositions are treated as identical, disputes about data appear to decide questions that are partly normative. An effective corpus report labels the precise proposition it is testing and avoids using the vocabulary of ordinary meaning as a shortcut around that choice.

The Judicial Debate in State v. Rasabout

In State v. Rasabout, 2015 UT 72, the Utah Supreme Court considered the allowable unit of prosecution for unlawful discharge of a firearm. The court affirmed the treatment of individual shots as discrete discharges. Associate Chief Justice Thomas R. Lee, concurring in part and in the judgment, advocated corpus-based examination of the contested expression and described how searchable language databases might test intuitions that dictionaries leave unresolved.

The important institutional debate concerned procedure as much as semantics. The opinion of the court objected to the introduction of an appellate judge's sua sponte corpus analysis without briefing and an opportunity for the parties to test it. Chief Justice Durrant, in a separate concurrence, was more receptive in principle while emphasizing caution and the usefulness of adversarial presentation. Those positions should not be compressed into the proposition that the Utah Supreme Court adopted corpus linguistics. The case documents contestation within the court over the evidentiary and institutional conditions of the technique.

A method can be analytically transparent yet procedurally unfair. If an appellate court constructs its own dataset, decides which contexts count, and produces a factual-seeming frequency after briefing has closed, the litigants may lack a meaningful chance to challenge collection bias, coding decisions, or semantic assumptions. In a criminal case these issues bear directly on notice and the fairness of the judicial process. Publishing queries in an opinion helps review, but transparency after judgment is not a substitute for an opportunity to contest evidence before the decision.

The decision supplies a methodological principle of general relevance: when corpus findings are offered in litigation, the procedure by which the court receives and scrutinizes them matters independently from their apparent empirical sophistication. Any practice importing linguistic science into an adversarial legal system must account for the role of parties, the court's permissible sources, and the timing of review.

Contemporary Judicial Applications

The Fourth Circuit's United States v. Boler, 115 F.4th 316 (4th Cir. 2024), involved the Sentencing Guidelines' use of loss and whether intended loss was encompassed. The published opinion includes a substantive separate discussion of corpus linguistics and a debate over the relationship between linguistic samples and regulatory context. The relevant lesson is not a universal holding about the definition of loss: the outcome depends on the Guidelines, their interpretive framework, and the respective judicial reasoning. A corpus result in that setting must be evaluated alongside the text and authority of the Guidelines themselves.

In Garibay v. Fox, No. CV-24-0091-PR (Ariz. Mar. 13, 2025), the Arizona Supreme Court addressed whether a constable's conduct met the statutory concept of misconduct and its relationship to immunity. The majority reported a search of the Corpus of Historical American English for 1900–1919, yielding 76 results, with 62 classified as intentional wrongdoing, one as interchangeable with negligence, and 13 indeterminate. These are figures reported by the court, not independently replicated counts in this monograph. The court used its linguistic analysis along with dictionaries and legal materials.

Chief Justice Ann A. Scott Timmer's separate concurrence in the result raised specific concerns about treating corpus analysis as a settled interpretive tool before adequate adversarial scrutiny. The published Arizona decision therefore contains both a worked judicial example and a caution about transparency, independent checking, and procedural safeguards. A reader should distinguish the court's holding from the majority's selected linguistic supports and from the concurrence's objection to methodology.

The growing judicial archive maintained by Brigham Young University lists opinions using corpus evidence across different subjects and procedural settings. Such a curated archive provides research leads; it is not by itself a representative census of all American courts or proof that a method commands nationwide judicial acceptance. The researcher should verify the operative opinion and identify whether corpus analysis appears in the majority, concurrence, dissent, expert submission, or merely a citation to academic work.

Together these decisions show the issue has matured beyond abstract academic proposal. Courts have examined specific concordance findings, but disagreement remains over sampling, legal context, interpretation of a statistic, and the authority of a judge to conduct independent empirical analysis. The method's actual institutional status is contingent on jurisdiction, court, record, and the kind of decision being made.

Legislative and Regulatory Corpora

A corpus of enacted statutes answers a different class of questions from a balanced corpus of public language. Legal terms may recur because drafters borrow statutory formulations, use established templates, or incorporate definitions from earlier acts. High frequency in legislation may therefore measure institutional drafting conventions rather than the ordinary public language of the period. The investigator should decide whether the disputed term is alleged to be a term of art, a general-language expression, or a formula whose recurrence itself is legally significant.

The Statutes at Large research interface supports retrieval of American federal enactments by official publication. Relevant metadata include enactment date, legal document identity, and version. A reproducible legislative corpus could then separate criminal law, taxation, administrative organization, and other statutory genres; it could trace phrase borrowing across versions and exclude duplicative codification text when estimating independent drafting choices. Those operations require actual text extraction and annotation before any quantitative claim can be made.

Gales and Solan's (2020) analysis of labor or service demonstrates why legislative phraseology may have a narrower technical range than surrounding general language. A phrase in a statute may also acquire meaning through judicial interpretation. A later judicial usage is evidence of doctrinal development, not necessarily of what the original legislature or general public understood. Mixing enactments, later judicial glosses, and commentary into one unmarked dataset creates circularity.

For regulations, the source chain can be even more intricate: consultation papers, proposed rules, final regulations, guidance, adjudications, and amended codes have distinct legal statuses. A linguistic study should classify these sources before pooling their vocabulary. If the interpretive question concerns a particular final rule, examples from an agency's policy commentary may illuminate register or purpose but cannot silently be treated as the authoritative words of the rule.

Expert Evidence and Adversarial Fairness

The procedural classification of corpus material matters. A lawyer's citation to an accessible dictionary, an expert linguist's report based on a proprietary corpus, an academic article submitted as authority, and a judge's independent search can enter legal reasoning through different routes. Questions of expert qualification, reproducibility, disclosure, and the opportunity to challenge methodology vary with the jurisdiction and procedural context. This monograph does not assert that one uniform evidence rule governs all corpus work.

In the United States, adversarial concerns were explicit in the Rasabout decision and resurfaced in the Garibay opinions. These courts show why a transparent dataset and the timing of its introduction are separate concerns. A party must have a fair opportunity to assess whether a corpus represents the claimed speech community and whether the selected examples instantiate the asserted sense.

When a qualified expert offers linguistic testimony, the court should examine the method used, its limits, and the relevant legal standard for receiving expert evidence. Where corpus work appears as judicial legal reasoning rather than trial evidence, the relevant legal and ethical constraints may be different. The correct answer is procedurally situated: provide enough method and source detail for meaningful response, identify any specially trained judgments, and avoid importing empirical conclusions into an appellate record without addressing the governing rules.

An adversarial process can improve research design. An opposing party might identify omitted genres, additional search forms, historically appropriate expressions, or an alternative coding of ambiguous contexts. A robust report anticipates these possibilities with a sensitivity appendix. The objective is to make disagreement inspectable, not to remove reasonable interpretive disagreement by statistical presentation alone.

Computational Tools, Automation, and Research Integrity

Digital corpus analysis increasingly uses tokenization, lemmatization, phrase matching, dependency parsing, named-entity recognition, semantic clustering, and large language models. Each method adds potential efficiency and potential error. Tokenization may split a legal compound; a lemma may merge historically distinct forms; a semantic embedding may reproduce training-corpus patterns while failing to respect a statute's technical definition. Automated coding can assist human researchers, but accuracy must be checked against the exact genre and time period at issue.

The most basic audit requires a traceable chain from source image or official text through extraction, correction, annotation, sampling, output table, and judicial or scholarly claim. An interface displaying an attractive frequency chart may hide OCR errors or a changing corpus index. A research packet should preserve the exact textual query, record failures, and describe preprocessing decisions. Where copyright or access restrictions prevent full-text redistribution, the research should still preserve stable identifiers and permitted reproducibility instructions.

Machine-assisted interpretation calls for special restraint in claims of precision. A language model may generate plausible usage examples that never appeared in historical sources. Such synthetic text is not valid evidence of attested historical language. A model may propose candidate senses and search variants, but each asserted instance must be tied to a source record, and each reported count must derive from an executed query. The distinction between generation and retrieval is an elementary requirement of source integrity.

Modern corpora also raise access inequalities. Some datasets are licensed and restrict search quotas or bulk downloads; others are built from digitized archival collections shaped by unequal preservation. Researchers should document these constraints and consider how they affect the population represented. Where a corpus cannot be shared, transparent metadata and independently checkable excerpts become particularly important for legal scrutiny.

Worked Research Protocols and Sensitivity Analysis

Consider a hypothetical dispute over whether a municipal rule regulating vehicles in a public garden includes an electric wheelchair. No actual ordinance or dataset is assumed. A corpus-only approach that samples everyday uses of vehicle may place cars, trucks, and buses at the center. A dictionary may include a broader range of conveyances. But a decision about the wheelchair would require examining the ordinance's definitions, accessibility provisions, purpose, statutory hierarchy, exemptions, and any applicable legal norms. The frequency of wheelchairs being called vehicles could be evidence of linguistic typicality, yet it would not exhaust the legal question.

An adequate research protocol would first quote and date the actual clause and identify the target interpretive question. It would then consider broad-language and legal-register corpora, articulate whether the relevant inquiry is a typical prototype or a category boundary, and choose search forms such as vehicle, vehicles, and contextually linked constructions. It would manually classify the relevant contexts while preserving uses that concern metaphorical vehicles, transport equipment, and mobility aids. A report would disclose if the corpus includes few references to wheelchairs, and it would refuse to treat rarity as semantic exclusion.

A second hypothetical concerns a historical statute prohibiting carrying arms in a specified setting. The investigator would identify the enactment, date, authentic wording, and disputed phrase; assemble period-appropriate genres; separate physical carrying, military deployment, metaphorical references, and disputed contexts; and run searches for morphological and syntactic alternatives. It would preserve the identities of republished documents and report whether a military sense dominates only in government or public genres. This is a research design example, not an empirical conclusion about constitutional language.

A third hypothetical concerns a multilingual EU instrument using parallel terms for professional services. The researcher would retrieve all relevant authentic versions and amendment records, compare terminological recurrence in the parallel corpus, and then apply the Court of Justice's rules concerning linguistic divergence, context, purpose, and EU-law concepts. Automatic sentence alignment could locate useful examples, but each claimed legal correspondence would be checked against authentic sources. The Commission's JRC-Acquis caveat would remain explicit in the report.

Sensitivity analysis belongs in all three protocols. Change the reference corpus and observe whether the classification proportions remain stable. Change the date interval, remove duplicated documents, increase the context window, treat uncertain examples separately, and test a broader as well as a narrower reading. Report which assumption drives any difference. If the legal conclusion is unaffected because the governing instrument defines the term expressly, the corpus research is still informative but its legal relevance is limited. The methodology gains credibility from showing where it adds little.

Evaluation of Competing Interpretive Positions

One position treats corpus inquiry as a promising discipline for controlling judicial intuition. It argues that examples selected from memory or isolated dictionaries are difficult to audit, whereas specified searches and transparent coding enable replication and criticism. Lee and Mouritsen's framework for ordinary-meaning inquiries makes this argument in concrete statutory settings. Its persuasive force is strongest when the contested proposition is empirically tractable, when the corpus closely matches the relevant context, and when the alternative method rests on unexplained intuitive assertions.

The conceptual objection has been articulated with particular force by Anya Bernstein. Her (2021) Half-Empirical Attitude argues that importing apparently ordinary instances from a literary or journalistic corpus into a legal dispute skips institutional questions about who is authorized to create law, how statutory speech acts operate, and which communicative conventions govern the specific legal genre. Her objection should be examined as a methodological challenge, not dismissed as resistance to measurement. If the target is legislative communicative meaning, the general corpus and a statute's enactment history may provide different kinds of information. If the target is the conventional sense of a commonplace word, broader public writing might nonetheless be probative. The researcher's burden is to justify that transition rather than to assume it.

Bernstein also asks what counts as data for an interpretive inquiry. A corpus may contain real language, but the mere authenticity of its texts does not establish that their distribution measures the question the court has asked. A sample of journalism may estimate the way journalists used a term, and a legal database may estimate the way lawyers used it. Neither, standing alone, necessarily estimates the recognition thresholds of ordinary readers or identifies a legal rule's scope. The methodological consequence is that an admissible corpus conclusion should contain a declared estimand: the precise linguistic quantity or property the investigator sought to estimate. An annotation manual, sampling frame, and sensitivity checks then become the means of testing whether the evidence answers that question.

A second position emphasizes the mismatch between observed typicality and legally relevant meaning. Tobia's experimental findings show why an empirical claim about prototypical use can differ from judgments about the outer boundary of a concept. This concern is particularly forceful in hard cases, because courts often decide disputes precisely where a category is unusual, an exception is important, or a statute reaches a circumstance that speakers rarely discuss.

John S. Ehrett's (2019) Against Corpus Linguistics offers a related but procedurally sharper objection. He questions the judicial use of corpora because different databases, coding choices, and opaque searches can compromise the hierarchy of legal sources and appellate review. The distinction between method choice and legal authority is crucial. A court can disclose an exact query while obscuring why that query is relevant. Conversely, a properly documented corpus study can clarify which claims depend on disputed selection assumptions, thereby making review more informed. The proper answer to this criticism varies with the judicial system's fact-finding conventions; a reproducibility file does not itself establish a litigant's right to be heard.

The criticism carries particular force where nonlawyers supply technical findings that judges then adopt as unreviewable premises. A protocol should therefore explain the identity and competence of the analyst, the creation of any annotations, the errors and disagreements in coding, and the procedural route by which opposing parties can test the analysis. The comparison is with whatever evidentiary and interpretive procedure applies to the actual court, not with an imagined universal standard of quantitative proof. A researcher publishing independently can disclose more methodological detail than the court's rules demand without confusing scholarly best practice with a binding evidentiary rule.

A third position emphasizes institutional procedure and methodological validation. The Utah and Arizona court debates show that judges can disagree about the admissibility, fairness, and expertise required to use quantitative linguistic material. Even an otherwise convincing result may be unsuitable as a decisive ground if the parties cannot challenge its production. The importance of process does not prove a method unreliable; it identifies conditions under which the legal institution can use the evidence fairly.

A fourth position situates corpus work within the entire legal interpretive framework. Theil's methodological account and the Court of Justice's multilingual reasoning underline that legal source authority, context, and institutional methods remain decisive in their respective settings. Corpus evidence can improve knowledge of a word's use while leaving the authoritative resolution to legal reasoning. This approach offers a way to combine empirical and doctrinal sophistication without treating them as competing substitutes.

The positions can be evaluated through a common question: what exact proposition has the method established, and what additional premises would make that proposition legally consequential? This question prevents a statistical count from carrying an unearned normative conclusion. It also prevents a critic from dismissing all empirical evidence merely because the law requires further judgment. The appropriate evidentiary weight is earned through relevance, source quality, inference validity, and procedural scrutiny.

Methodological Findings and Authorial Synthesis

The literature and case studies support several interconnected conclusions. The first concerns specification. Ordinary meaning is an incomplete research instruction unless it identifies the text, context, population, period, and aspect of meaning at issue. A question about the typical use of a word, its possible extension, and its legal term-of-art character can demand three different research designs. The analyst should specify which one is being pursued before collecting examples.

The second concerns source fit. Representative language depends on the target population, not just the size or institutional prestige of a dataset. A modern balanced-language corpus cannot stand in for a historical community merely by offering millions of words. A parallel EU corpus cannot displace authentic versions of legislation. A court-opinion corpus may reveal judicial drafting conventions but overrepresent terms propagated through precedent. These conclusions follow from ordinary principles of sampling, source criticism, and the legal status of texts, applied together.

The third concerns interpretive inference. The proper result of a corpus analysis is a bounded, inspectable proposition: the specified sample, collected and classified under described rules, exhibits a given pattern. Any proposition about legal coverage requires an account of why that pattern is pertinent under the applicable law. The nonappearance fallacy shows the danger of passing from a negative search result to a categorical semantic exclusion, while the judicial cases show the equally important danger of treating a few selected examples as conclusive.

The fourth concerns procedural justice. Reproducibility is necessary for contestability but does not by itself provide a litigant with an opportunity to respond. The experiences in Rasabout and Garibay show that method transparency, expert competence, and adversarial testing require explicit institutional attention. A judge's independent data analysis may be better documented than a dictionary intuition, but its procedural legitimacy must still be established in the relevant forum.

The fifth concerns research honesty. This monograph has inspected published cases, institutional documentation, and scholarly arguments, and it has compared their methods and implications. It has not carried out independent corpus searches or new participant experiments. Therefore it claims no original usage frequencies, statistical estimates, jurisdiction-wide judicial adoption rates, or discovery of a new linguistic regularity. Its authorial contribution is the systematic mapping of the distinct evidentiary steps and the evaluation of how they interact, together with transparent illustrative research protocols.

These findings yield a practical rule for scholarly writing: report what the language records show, state why those records are informative, explain what they leave unresolved, and then conduct the legal analysis under the proper authority. That sequence gives empirical evidence a meaningful role and makes the residual legal judgment visible. It also gives future researchers a testable agenda: compare alternative corpus designs on the same disputed expressions, publish independent annotation results, and investigate whether different methods converge when the relevant linguistic target is specified in advance.

Conclusion

Legal corpus linguistics has become a substantial part of contemporary scholarship on language and law and has entered identifiable judicial decisions. Its strongest contribution is to make claims about usage open to systematic examination. Its central vulnerability is a mismatch between sampled uses and the legally significant meaning that a court seeks to determine. Large datasets, computer-generated counts, and frequency tables do not dissolve that problem. They create useful evidence when the population, genre, date, legal question, and inferential steps are properly specified.

The case studies demonstrate both the opportunity and the limits. Muscarello and Taniguchi present contested ordinary-language problems later reanalyzed with corpus methods; Rasabout made the debate over judicial initiative and adversarial fairness explicit; Boler displays tension between linguistic sampling and the interpretive structure of federal sentencing materials; and Garibay offers reported historical corpus counts alongside a concurrence demanding scrutiny. The EU materials establish that parallel language resources must be interpreted with respect to multilingual legal authority and doctrinal context.

A mature practice therefore treats the corpus as a carefully designed evidentiary instrument, its output as a qualified observation, and the legal conclusion as a further act of reasoned interpretation. The method is most credible when its search choices are documented, alternative analyses remain possible, source authority is preserved, and its limitations appear directly beside its conclusions. The future of this field depends on more exact research design, replication, historical and multilingual competence, and procedures that allow empirical claims to be contested before they influence legal outcomes.

References

  • Bernstein, Anya. 2021. “Legal Corpus Linguistics and the Half-Empirical Attitude.” Cornell Law Review 106: 1397–1456. Cornell Law Review publication.
  • Bernstein, Anya. 2021. “What Counts as Data?” Brooklyn Law Review 86: 435–459. Journal article record.
  • Ehrett, John S. 2019. “Against Corpus Linguistics.” Georgetown Law Journal Online 108: 51–73. Georgetown Law publication.
  • Gales, Tammy, and Lawrence M. Solan. 2020. “Revisiting a Classic Problem in Statutory Interpretation: Is a Minister a Laborer?” Georgia State University Law Review 36: 491. University journal record.
  • Lee, Thomas R., and Stephen C. Mouritsen. 2018. “Judging Ordinary Meaning.” Yale Law Journal 127: 788–879. Journal full-text record.
  • Lee, Thomas R., and Stephen C. Mouritsen. 2021. “The Corpus and the Critics.” University of Chicago Law Review 88 (2): 275. Journal record.
  • Mouritsen, Stephen C. 2017. “Corpus Linguistics in Legal Interpretation—An Evolving Interpretative Framework.” International Journal of Language & Law 6: 67–89. Journal full text.
  • Theil, Stefan. 2025. “Carefully Tailored: Doctrinal Methods and Empirical Contributions.” Oxford Journal of Legal Studies 45 (4): 1047–1075. Publisher DOI record.
  • Tobia, Kevin P. 2020. “Testing Ordinary Meaning.” Harvard Law Review 134: 726. Journal publication.
  • Tobia, Kevin P. 2021. “The Corpus and the Courts.” University of Chicago Law Review Online (March 5). Journal essay.
  • Wright, David. 2026. “Corpus Perspectives on Legal Discourse.” International Journal of Corpus Linguistics 31 (1): 1–12. Publisher DOI record.
  • Muscarello v. United States, 524 U.S. 125 (U.S. Supreme Court 1998). Judicial opinion.
  • Taniguchi v. Kan Pacific Saipan, Ltd., 566 U.S. 560 (U.S. Supreme Court 2012). Judicial opinion.
  • State v. Rasabout, 2015 UT 72 (Utah Supreme Court 2015). Judicial opinion.
  • United States v. Boler, 115 F.4th 316 (U.S. Court of Appeals for the Fourth Circuit 2024). Judicial opinion.
  • Garibay v. Fox, No. CV-24-0091-PR (Arizona Supreme Court, March 13, 2025). Judicial opinion.
  • Consorzio Italian Management and Catania Multiservizi SpA v. Rete Ferroviaria Italiana SpA, Case C-561/19, ECLI:EU:C:2021:799 (Court of Justice of the European Union 2021). Authentic EU-law judgment record.
  • U.S. Government Publishing Office. n.d. “Statutes at Large.” Official collection guide.
  • U.S. Government Publishing Office. n.d. “United States Statutes at Large.” Official collection.
  • English-Corpora.org. n.d. “Corpus of Contemporary American English: Texts.” Corpus composition documentation.
  • English-Corpora.org. n.d. “Corpus of Historical American English: Texts.” Corpus composition documentation.
  • European Commission, Joint Research Centre. n.d. “JRC-Acquis.” Project documentation and restrictions.
  • Brigham Young University Law School. n.d. “Legal Opinions.” Law and Corpus Linguistics. Curated decision index.
  • Sources of Language and Law (SOULL). n.d. “Data Collections.” Corpus catalog.