⏳ Historical linguistics & language change

A survey · about 17 minutes to read · every paper named here has its own page in this library

Historical linguistics studies how languages change over time and how they are related to one another. Its questions are old ones: why Latin became French, Spanish and Romanian; why English gay, virus and cell mean something different today from what they meant two centuries ago; how a noun meaning "property" can turn into a grammatical marker meaning "of". The field reconstructs lost ancestor languages, traces the history of individual words (etymology), and describes the regular paths along which sounds, forms, constructions and meanings shift.

The papers on this page come from two traditions that have only recently begun to talk to each other. One is philological and comparative: close readings of Biblical Hebrew and Aramaic, Proto-Semitic reconstruction, Old and Middle English syntax, the grammaticalization of German modals and North Germanic articles. The other is computational: since about 2010, researchers in natural language processing have used large dated corpora and word embeddings to detect and measure semantic change automatically, built shared benchmarks for the task, and started to parse, tag and lemmatize historical texts with the same tools used for modern ones.

Both share one problem: evidence. Records are patchy and skewed toward certain genres, and old texts often reach us through scribes who changed them. Much of the work here asks how to draw sound conclusions from such material.

What it is

Historical linguistics (also called diachronic linguistics) contrasts with synchronic linguistics, which describes a language at a single point in time. The distinction goes back to Ferdinand de Saussure's lectures in the early twentieth century, but the discipline itself is older, rooted in nineteenth-century comparative philology.

The core notions are few but powerful:

• Sound change is the gradual, typically regular modification of pronunciation. Grimm's law (Latin pater, English father) is the classic example; regularity makes the comparative method work.
• Cognates are words in different languages that descend from a common ancestor form; loanwords are borrowed from one language into another; false friends look alike but differ in meaning. Telling these apart is the first step in establishing relationships between languages.
• Reconstruction recovers unattested ancestor forms, marked with an asterisk (Proto-Semitic *-um, *-im, *-am). The comparative method works across related languages; internal reconstruction works from irregularities inside a single language.
• Semantic change covers broadening, narrowing, metaphor, metonymy, amelioration and pejoration of word meaning. Computational work usually calls it lexical semantic change (LSC) and treats its detection as a task (LSCD).
• Grammaticalization is the process by which lexical items become grammatical markers, often losing phonetic substance and agreement along the way: a demonstrative becomes a definite article, a noun for "thing" becomes a genitive particle, a modal verb becomes a conditional marker.

What sets the topic apart from its neighbours is the explicit time dimension: the object of study is not a meaning or a form, but the difference between two or more stages.

History and milestones

Comparative philology

Modern historical linguistics began around 1800, when William Jones, Rasmus Rask, Franz Bopp and Jacob Grimm recognised systematic correspondences between Sanskrit, Greek, Latin and Germanic. In the 1870s the Neogrammarians in Leipzig formulated the principle that sound laws operate without exceptions, with analogy accounting for apparent irregularities. Semitic studies developed in parallel, and work such as The Reconstruction of the Proto-Semitic Genitive Ending and a Suggestion on its Origin continues that tradition today.

Twentieth-century developments

The twentieth century added sociolinguistic studies of change in progress (William Labov, from the 1960s), grammaticalization theory (Heine, Hopper, Traugott and others), and generative accounts of syntactic change through reanalysis by learners. Morris Swadesh's basic vocabulary lists of the 1950s remain the usual sample for cognate studies. Toward the end of the century came historical treebanks: the Penn-Helsinki corpora of Middle English and, in the early 2000s, the YCOE treebank of Old English prose, which made quantitative syntactic study of earlier English feasible.

The computational turn

Early computational work used historical relationships as a tool rather than a subject. The oldest and one of the most cited papers here, Using cognates to align sentences in bilingual corpora from 1992, exploited the fact that English and French share many word beginnings to align parallel texts for machine translation.

The decisive change came with very large dated corpora. The Google Books Ngram data (released in 2010) and the Corpus of Historical American English (COHA) gave researchers decade-by-decade word statistics over two centuries. In 2011 A distributional similarity approach to the detection of semantic change in the Google Books Ngram corpus. showed that count-based context vectors could flag words whose meaning shifted. After word2vec (2013), diachronic embeddings became the standard instrument, culminating in the two Stanford papers of 2016 that proposed statistical laws of semantic change.

A second phase, from 2018, concentrated on evaluation: the DURel annotation framework, the 2018 survey from Oslo, and the SemEval-2020 shared task. A dedicated workshop series, LChange, began in 2019; several of the newest papers here come from its sixth meeting in 2026. A third phase replaced static vectors with contextual models such as BERT, and most recently with large language models that generate definitions or reason about types of change.

Most papers here appeared in 2018 or later, helped by the LChange workshops and a wave of new historical corpora at LREC 2026.

Main problems and tasks

Detecting and measuring semantic change

The best-developed computational task asks: given two corpora from different periods, which words changed meaning, and by how much? SemEval-2020 split this into a binary decision (did a word gain or lose a sense?) and a graded ranking.

Characterising change

Recent work also asks how a word changed, sorting shifts by type (broadening, narrowing, metaphor, metonymy, amelioration, pejoration). Semantic Change Characterization with LLMs using Rhetorics and Interpretable Word Sense Representations via Definition Generation: The Case of Semantic Change Analysis are examples. A related distinction, between culturally driven change (new technology, new referents) and regular linguistic drift, is drawn in Cultural Shift or Linguistic Drift? Comparing Two Computational Measures of Semantic Change.

Cognate detection and reconstruction

Identifying cognate sets across related languages, recovering regular sound correspondences, and proposing proto-forms are the classic tasks of the comparative method. Computational tools now automate parts of them, as in Automated Cognate Detection as a Supervised Link Prediction Task with Cognate Transformer and Cognate Detection for Historical Language Reconstruction of Proto-Sabean Languages: the Case of Ge'ez, Tigrinya, and Amharic. Philological reconstruction continues in parallel, for example of the Hebrew participle in The Biblical Hebrew Feminine Singular Qal Participle: A Historical Reconstruction.

Processing historical text

Before historical text can be studied quantitatively it must be tagged, lemmatized and parsed, yet old language stages have little annotated data and highly variable spelling. Common strategies are projecting tags from modern translations, as in Part-of-Speech Tagging for Middle English through Alignment and Projection of Parallel Diachronic Texts, transferring parsers from related languages, as in Diachronic proximity vs. data sparsity in cross-lingual parser projection. A case study on Germanic, and converting old treebanks to Universal Dependencies.

Tracing syntactic and grammatical change

How did object pronouns move between Old and Middle English, how did articles grow out of demonstratives, did scientific prose become more compact, and can automatic parsers measure such trends at all?

Dating texts and textual transmission

In Hebrew Bible studies, linguistic features have long been used to date books as "early" or "late". Several papers here question whether that is possible when manuscripts differ word by word, and argue that language history cannot be separated from textual criticism.

Methods and schools

Philological and comparative

Benjamin Suchard uses Biblical Aramaic as a control for Hebrew: a sound change found in both, read within one shared reading tradition, likely belongs to that tradition rather than to the spoken language. Aaron Rubin, in Notes on the Genitive Exponents of Some Modern Arabic Dialects, groups a dozen dialect possessive markers by just two kinds of source. Wilfred Watson's Terms for Occupations, Professions and Social Classes in Ugaritic: An Etymological Study sorts Ugaritic job titles into inherited words and loans.

Formal and corpus-based syntax

Theoretical syntacticians working with historical corpora appear here through papers in Glossa: the comparative constructions of German in The syntax of comparison constructions in diachronic and dialectal perspective, conditional sollte in Should a conditional marker arise … The diachronic development of conditional sollte in German, pronoun placement in On the syntax of object pronouns in Old English and Early Middle English, the rise of the article in Familiar vs. unique in a diachronic perspective. Case study of the rise of the definite article in North Germanic, and prosody recovered from verse in Metre and clitics in Old English and Old Saxon.

Count-based distributional methods

The first computational approaches compared a word's co-occurrence counts across periods over one shared set of contexts, so no alignment was needed. Gulordava and Baroni's 2011 work is the classic example here; the interpretable dependency-slot profiles of Transparent Semantic Change Detection with Dependency-Based Profiles revive the idea with syntactic contexts.

Static diachronic embeddings

From 2014 onward, researchers trained separate embedding models per time slice and aligned them, most often with orthogonal Procrustes, or initialised each slice from the previous one. Change is measured as cosine distance between a word's aligned vectors, or as a shift in its nearest neighbours. The HistWords resources released with Diachronic Word Embeddings Reveal Statistical Laws of Semantic Change became a common starting point.

Contextual and usage-based models

Contextual models such as BERT give each occurrence of a word its own vector. Change can then be measured by clustering occurrences into usage types and comparing their distributions across periods, or by averaging pairwise distances between usages. Analysing Lexical Semantic Change with Contextualised Word Representations set out this approach; Using BERT to Explore Lexical Semantic Change of Prepositions applies it to a word class usually neglected. The current view, reflected in The LSCD Benchmark: a Testbed for Diachronic Word Meaning Tasks, breaks the task into three stages: judging whether two uses mean the same (Word-in-Context), inducing senses by clustering, and comparing sense distributions.

Generative models

Large language models now enter the field in several roles: generating definitions as readable sense labels, reasoning about the kind of change with rhetorical tests such as zeugma, drafting historical treebank sentences that humans then correct, and even generating new Old English text, as in AI-Driven Generation of Old English: A Framework for Low-Resource Languages. Generated output still needs human checking.

Main debates

• Static versus contextual. Contextual embeddings seem better suited to sense-level change, yet in SemEval-2020 type-based static embeddings won. Later benchmarks with better word-in-context models have narrowed or reversed that gap.
• Frequency confounds. Many change scores correlate strongly with word frequency, so apparent laws or rankings may partly reflect frequency artefacts. Controls for frequency and polysemy are now expected.
• Interpretability. Embedding distances show that a word moved, not why; definition generation and dependency profiles respond to this.
• Parser reliability. Whether trends computed from automatic parses of old text are real or artefacts of the parser is examined directly in Syntactic Language Change in English and German: Metrics, Parsers, and Convergences.

Datasets, resources and benchmarks

Diachronic corpora. The Google Books Ngram data and COHA dominate English-language work. Germanic historical text comes from the Deutsches Textarchiv (DTA) and the DWDS corpora; Swedish from Kubhist; Latin from LatinISE. The Royal Society Corpus, covering scientific writing from 1665 to the late twentieth century, supports a cluster of papers on scientific English. New resources here include an Urdu corpus of printed books from 1800 to 1950 with page images (Scripting History: A Diachronic Urdu Text and Image Corpus from the 18Th to 19Th Centuries), a multimodal corpus of Bundestag speeches by long-serving deputies (Voices across Decades: A Multimodal Diachronic Corpus of German Bundestag Debates (GerParlDia-MM)).

Historical treebanks. YCOE for Old English and the Penn-Helsinki corpora for Middle English are the main syntactic resources; papers here convert them toward Universal Dependencies and enrich them with lemmas.

Semantic change benchmarks. DURel, the SemEval-2020 Task 1 data (English, German, Latin, Swedish), the DWUG family of word usage graphs, and language-specific sets such as NorDiaChange, RuShiftEval and ChiWUG. AmDi adds sense-labelled German data across decades: AmDi - Ambiguous Words Diachronic Dataset.

Etymological resources. Etymological Wordnet: Tracing The History of Words mines word-origin links from Wiktionary; EtymoLink: A Structured English Etymology Dataset structures the Online Etymology Dictionary. Cognate data include the Arabic–Hebrew set in When Similar Means Different: Evaluating LLMs on Arabic–Hebrew Cognates.

Lexical resources for older stages. Towards the Construction of a WordNet for Old English plans an Old English wordnet.

Landmark papers in this library

Hamilton, Leskovec and Jurafsky: statistical laws (2016)

Diachronic Word Embeddings Reveal Statistical Laws of Semantic Change is the most cited work on this page. The authors trained decade-level embeddings for English, French, German and Chinese, aligned them with Procrustes, and checked their measurements against known historical shifts. They proposed two regularities: common words change meaning more slowly (a "law of conformity"), and, frequency held equal, words used in more varied contexts change faster (a "law of innovation"). The released HistWords embeddings shaped a decade of follow-up work, and later replications showed how sensitive such laws are to frequency effects.

Simard, Foster and Isabelle: cognates for alignment (1992)

Using cognates to align sentences in bilingual corpora used a crude, dictionary-free notion of cognate (shared four-letter prefixes, identical numbers and punctuation) to improve sentence alignment of the Canadian Hansards. It shows historical relatedness between languages serving as a practical engineering signal.

The SemEval-2020 shared task (2020)

SemEval-2020 Task 1: Unsupervised Lexical Semantic Change Detection gave the field its first common test across four languages, with gold labels from large-scale human judgments of usage pairs. Dozens of teams took part; neither subtask was solved, and the surprising strength of static embeddings, together with their sensitivity to frequency, set the agenda for the following years.

TimeLMs (2022)

TimeLMs: Diachronic Language Models from Twitter released a series of Twitter language models refreshed every quarter, and showed that models degrade on text from later periods and that continued training softens this decline.

Cultural shift versus linguistic drift (2016)

Cultural Shift or Linguistic Drift? Comparing Two Computational Measures of Semantic Change compared a global measure (how far a word's vector moves) with a local one (how its neighbourhood changes). The local measure picks up culturally driven noun changes such as gay and virus; the global one picks up regular processes such as subjectification in actually or must.

The 2018 survey

Diachronic word embeddings and semantic shifts: a survey brought together work scattered across NLP, information retrieval and political science, linked it to historical linguistics, and listed open problems: languages other than English, small datasets, reliable gold standards, and sense-aware analysis. It became the standard entry point for newcomers.

Gulordava and Baroni (2011)

A distributional similarity approach to the detection of semantic change in the Google Books Ngram corpus. compared words between the early 1960s and the late 1990s using bigram counts, and evaluated against human ratings rather than a handful of anecdotes. It showed that context similarity tracks perceived change better than frequency change,.

Contextualised representations (2020)

Analysing Lexical Semantic Change with Contextualised Word Representations was among the first to apply BERT to the problem without any sense inventory. Clustering usage vectors let the authors trace narrowing (coach) and broadening (disk).

DURel (2018)

Diachronic Usage Relatedness (DURel): A Framework for the Annotation of Lexical Semantic Change solved an annotation problem: instead of asking people to define senses, it asks how related two uses of a word are, on a four-point scale, and compares pairs within and across periods. The method is language-independent and underlies SemEval-2020 and the later DWUG datasets.

Definition generation (2023)

Interpretable Word Sense Representations via Definition Generation: The Case of Semantic Change Analysis generated a short dictionary-style definition for each usage and used them to label sense clusters, so that a shift in usage can be read as a shift in definitions.

Sound changes in the reading tradition (2019)

Sound Changes in the (Pre-)Masoretic Reading Tradition and the Original Pronunciation of Biblical Aramaic separated sound changes of the shared Tiberian reading tradition from older ones of spoken Hebrew, giving a relative chronology and a reconstruction of Biblical Aramaic pronunciation.

The Qumran Judges scrolls (2013)

The Qumran Scrolls of the Book of Judges: Literary Formation, Textual Criticism, and Historical Linguistics compared the Masoretic text of Judges with Qumran fragments and found that late additions are written in "classical" Hebrew, with early and late features scattered unpredictably across witnesses. It is a central text in the argument that linguistic profiles alone cannot date biblical books.

Citations concentrate in papers from 2015 to 2019; the newest work has not yet had time to be cited.

People

Dominik Schlechtweg, now at the University of Stuttgart and the University of Oslo, is the most prolific author on this topic. His work runs from the DURel annotation framework through the SemEval-2020 shared task to the LSCD Benchmark, and has done much to give semantic change detection reliable evaluation.

William L. Hamilton, with Jure Leskovec and Dan Jurafsky at Stanford University, wrote the two 2016 papers that made diachronic embeddings a mainstream method and proposed the laws of conformity and innovation.

Andrey Kutuzov of the University of Oslo led the 2018 survey with Lilja Øvrelid and Erik Velldal, and later co-authored the work on definition generation.

Mario Giulianelli and Raquel Fernández, associated with the University of Amsterdam, introduced usage-based analysis of semantic change with BERT and then the definition-based sense labels.

Sabine Schulte im Walde of the University of Stuttgart co-designed DURel and, with Filip Miletić, studied how English noun compounds lose compositionality over time in Modeling the Evolution of English Noun Compounds with Feature-Rich Diachronic Compositionality Prediction.

Kristina Gulordava and Marco Baroni wrote one of the founding papers of distributional change detection; Baroni is now at ICREA and Universitat Pompeu Fabra.

Gerard de Melo, at the Hasso Plattner Institute and the University of Potsdam, built the Etymological Wordnet.

Benjamin D. Suchard, of Leiden University and Ghent University, works on Semitic historical phonology and morphology, from the Tiberian reading tradition to the Proto-Semitic genitive.

Robert Rezetko of the University of Copenhagen is a leading critic of linguistic dating of the Hebrew Bible; besides the Judges study he wrote the review Currents in the Historical Linguistics and Linguistic Dating of the Hebrew Bible: Report on the State of Research as Reflected in Recent Major Publications.

Elke Teich and Marie Pauline Krielke of Saarland University study the history of scientific English and German with parsed corpora and information-theoretic measures, together with Diego Alves and Stefania Degaetano-Ortlieb.

Anne Breitbarth of Ghent University works on Germanic historical syntax, from conditional sollte to parser-based measurement of syntactic change.

Luca Brigada Villa of the University of Pavia and Martina Giarda bring Old English into Universal Dependencies, converting YCOE and testing transfer from modern Germanic languages.

Sanja Štajner, with Ruslan Mitkov, measured twentieth-century change in readability and text complexity across the Brown family corpora.

Michel Simard, George F. Foster and Pierre Isabelle (Isabelle now at the National Research Council Canada) wrote the 1992 cognate alignment paper.

Institutions and places

The computational side of the topic is strongly European. The University of Oslo and the University of Stuttgart are the main centres for semantic change benchmarks, joined by the University of Amsterdam, University of Zurich and ETH Zurich. Saarland University leads work on the history of scientific language. Ghent University and Leiden University contribute Germanic syntax and Semitic historical linguistics; the University of Copenhagen is the home of critical work on Hebrew dating. Other homes of authors here include the University of Cambridge, University College London, Stanford University, Ludwig-Maximilians-Universität München, the University of Mannheim, the University of Potsdam and the University of Pavia.

Germany leads by country, followed by the United Kingdom, the United States, the Netherlands, Belgium and Norway.

The venues reflect the two traditions. Glossa: a journal of general linguistics carries most of the theoretical work on historical syntax; LREC proceedings carry new corpora and resources; the LChange workshops collect computational semantic change research; and journals such as the Journal of Hebrew Scriptures, Folia Orientalia and the Journal of Semitic Studies carry Semitic philology.

Open questions

• What exactly is being measured? Change scores mix sense change, shifts in topic or context, and changes in frequency. Separating these, and deciding which counts as "semantic change", remains unsettled.
• Beyond detection. Automatic methods can rank words by degree of change, but explaining the type of change, its cause and its timing is still largely manual work.
• Small and noisy data. Most methods need large corpora; older stages of most languages offer little text, inconsistent spelling and OCR noise.
• Languages beyond English and German. Benchmarks now exist for several European languages, Russian and Chinese, but coverage of the world's languages is thin. New resources for Urdu, Ethio-Semitic and Romance are first steps.
• Reliability of automatic analysis. Parser-dependent results, tagger accuracy across genres (as in Cross-Genre Learning for Old English Poetry POS Tagging) and the quality of machine-generated historical text all need validation before conclusions about language history can rest on them.
• Cognates versus loans. Neural detectors learn sound correspondences but still confuse inherited words with borrowings.
• Continuity in grammar. How a grammar with discrete rules produces apparently gradual change, raised in Detecting Syntactic Change Using a Neural Part-of-Speech Tagger, is an old question that new models have not answered.
• Dating texts. Whether linguistic features can date texts whose manuscripts kept changing during transmission is contested, as Diachronic Study of the Hebrew Bible: A Field in Crisis and Historical linguistics and linguistic dating of the Hebrew Bible 2015–2018: The Rezetko and Young assessment of the status quaestionis show from different sides.

Neighbouring topics

Much of this topic lies on the Linguistics & languages shelf, with a smaller computational part on Natural language processing and an early link to Machine translation.

The largest overlap is with Semitic languages and Hebrew & Aramaic, where reconstruction, the Masoretic reading tradition and the linguistic dating of biblical books are long-standing concerns; a small overlap with Akkadian and cuneiform NLP reflects the place of Akkadian in Semitic reconstruction. Lexical semantics & word embeddings supplies the representations behind semantic change detection, and Language models & pretraining the contextual and generative models that now dominate it.

Historical syntax draws on Universal Dependencies, Treebanks, Part-of-speech tagging and Syntax & grammar formalisms; grammaticalization touches Morphology and Idioms & multiword expressions; diachronic narrative and presupposition studies touch Discourse analysis and Implicature. Cognates link to Multilingual & cross-lingual NLP, and most historical stages count as Low-resource languages. See also Latin, Ancient Greek and Evaluation & benchmarks.

Key works in this library

• Diachronic Word Embeddings Reveal Statistical Laws of Semantic Change — aligned decade embeddings in four languages and two proposed laws of change.
• Using cognates to align sentences in bilingual corpora — shared word beginnings as a cheap signal for aligning parallel text.
• SemEval-2020 Task 1: Unsupervised Lexical Semantic Change Detection — the first multilingual shared benchmark for change detection.
• TimeLMs: Diachronic Language Models from Twitter — language models refreshed every quarter to keep up with change.
• Diachronic word embeddings and semantic shifts: a survey — the standard overview of the computational field.
• A distributional similarity approach to the detection of semantic change in the Google Books Ngram corpus. — an early count-based method evaluated against human ratings.
• Analysing Lexical Semantic Change with Contextualised Word Representations — BERT usage clusters for tracing narrowing and broadening.
• Diachronic Usage Relatedness (DURel): A Framework for the Annotation of Lexical Semantic Change — rating use pairs instead of defining senses.
• Interpretable Word Sense Representations via Definition Generation: The Case of Semantic Change Analysis — readable sense labels from generated definitions.
• The LSCD Benchmark: a Testbed for Diachronic Word Meaning Tasks — one repository for change detection and its subtasks.
• Using Modern Languages to Parse Ancient Ones: a Test on Old English — Icelandic data helps parse Old English.
• Diachronic Analysis of Multi-word Expression Functional Categories in Scientific English — stance expressions gain ground in scientific writing.
• Sound Changes in the (Pre-)Masoretic Reading Tradition and the Original Pronunciation of Biblical Aramaic — Aramaic as a control for dating Hebrew sound changes.
• The Qumran Scrolls of the Book of Judges: Literary Formation, Textual Criticism, and Historical Linguistics — manuscript variation undermines linguistic dating.
• Automated Cognate Detection as a Supervised Link Prediction Task with Cognate Transformer — cognate sets predicted from aligned word lists.