🏛️ History

A survey · about 11 minutes to read · every paper named here has its own page in this library

The History shelf, a subfield of the arts and humanities, collects work in which computational methods serve the study of the past and of culture: old spellings and dead languages, scripture and its translations, inscriptions, rabbinic and medieval documents, literary style and the history of ideas. It is the computational wing of what is now called the digital humanities, and many papers here also stand on linguistics, because the raw material of the humanities is, above all, text.

The papers share a set of practical problems. Historical texts are written in languages or varieties for which modern tools were never trained; their spelling is unstable; their surviving copies are damaged, abbreviated or noisily digitised; and annotated data is scarce. At the same time, the questions asked of them are scholarly rather than commercial: who wrote a passage, how a word changed its meaning, which source a compiler used, what a translator added. The shelf therefore mixes engineering, such as parsers, taggers and language models for ancient languages, with analysis that answers questions from history, philology and literary studies.

What it is

The shelf covers three overlapping areas.

• Language technology for historical and ancient languages: normalisation of old spelling, lemmatisation, part-of-speech tagging, morphological analysis, dependency parsing, named entity recognition and language models for Latin, Ancient Greek, Biblical and Rabbinic Hebrew, Aramaic, Syriac, Classical Armenian, Old Georgian, Old High German and Classical Japanese.
• Digital philology and heritage infrastructure: annotation platforms for inscriptions, encoding of epigraphic records in TEI, controlled vocabularies, lexicons built from parallel texts, and corpus interfaces for scholars who do not write query languages.
• Computational analysis of culture: authorship and source hypotheses, stylistic variation, translation behaviour, metaphor, and conceptual change in science.

It differs from its large neighbour, Language and Linguistics, in emphasis rather than method: the goal is usually a statement about texts, authors or periods, not about language as such. Its closest topics are Hebrew & Aramaic, Latin, Ancient Greek and Historical linguistics & language change.

History and milestones

Humanities computing is much older than the papers collected here. Roberto Busa's index of the works of Thomas Aquinas, begun with IBM in 1949, is usually named as its starting point, and concordances, stylometry and textual encoding standards such as the Text Encoding Initiative (from 1987) shaped the field for decades. This shelf records a later phase, in which statistical and then neural NLP were brought to bear on historical material.

Rules and alignments (2011–2014). The earliest works here deal with noisy and non-standard text. Rule-Based Normalization of Historical Texts learned spelling rewrite rules for Early New High German from two aligned editions of the Luther Bible. A study of Hebrew-script documents from the Cairo Genizah separated Hebrew, Aramaic and Judeo-Arabic in noisy OCR output, and a Frankfurt project on parallel Bibles in older Germanic used verse alignment to link Old High German and Old Saxon to their relatives.

Statistical and neural tools (2018–2022). Spelling normalisation was revisited for Early English letters, comparing rules, edit distance and machine translation. Neural lemmatisers and taggers reached Classical Armenian, Old Georgian and Syriac. Historical pre-training arrived with MacBERTh, a BERT model trained on five centuries of English. The first LT4HALA workshop, devoted to language technology for historical and ancient languages, met in 2020, giving the area a regular venue.

Language models and LLMs (2024–2026). Most papers on the shelf are from 2026. They adapt pretrained models to Ancient Hebrew and Rabbinic Hebrew, use large language models to correct Latin parses or extract facts from biographies, and ask whether such models behave in line with humanities theories.

Main problems and tasks

Normalisation and basic annotation

Before a historical text can be searched or parsed, its words must be matched to standard forms and lemmas. Spelling normalisation, lemmatisation and tagging recur throughout the shelf, from Luther-era German to the Justinian's Digest. Prefixes and clitics are a special difficulty in Semitic languages: a single written word may contain a conjunction, a preposition and an article. See Morphology and Part-of-speech tagging.

Morphology and syntax of ancient languages

Finite-state analysers for Ancient Greek and Hebrew showed how to model accent and stress shifts. Latin dependency parsing was the subject of the EvaLatin 2026 shared task, with entries from Uppsala and the retrieval-augmented THIVLVC system; both found poetry much harder than prose. These connect to Dependency parsing and Universal Dependencies.

Named entities and information extraction

Finding people, places and citations in old documents is central to historical research. Papers here probe Latin named entities without training data, train recognisers for medieval French charters from small entity lists, and tag citations in medieval Hebrew responsa. See Named entity recognition and Information extraction.

Restoration, lexicons and senses

Damaged documents invite text restoration, as in the infilling of gaps in medieval Latin notarial deeds. Lexical resources include an Aramaic–Hebrew translation lexicon built from the Targums, a thesaurus of Biblical Hebrew, and work on WordNet synsets for Ancient Greek. Word sense induction separates the meanings of ambiguous Biblical Hebrew words without a large model; this links to Word sense disambiguation.

Authorship, style and sources

Classic questions of textual scholarship reappear as classification tasks. One paper recasts the Documentary Hypothesis of the Torah as span classification; another searches a medieval anthology for lost Midrash Tanhuma material and had an expert confirm most of the candidates. Newer work builds a content-controlled authorship corpus from Wikipedia and uses stylometry to tell human French fiction from imitations generated by language models.

Translation, meaning and conceptual change

Interlinear glosses of the New Testament serve as a neutral baseline for measuring how far translators depart from the source. Greek–Latin parallel sentence mining uses Boethius' translations of Aristotle. Metaphor detection reaches Classical Japanese and German sermons, and a network model traces the Chemical Revolution as oxygen displaced phlogiston in the Royal Society's texts.

Methods and schools

The shelf shows clearly how methods for historical text have shifted while the problems stayed the same.

Rules and resources. Hand-built or learned rewrite rules, finite-state morphology, dictionaries and lexicon constraints remain important because data is scarce and scholars want transparent decisions. A study of Emergent Modern Hebrew diacritisation found that restricting outputs to attested vocalisations gave larger gains than a bigger model.

Statistical learning and alignment. Parallel texts, especially Bible translations and Targums, provide free supervision: aligned verses yield rewrite rules, translation lexicons and projected syntactic trees.

Transfer and pre-training. A central debate is whether to adapt a modern model or train one on historical data from scratch. MacBERTh argued for training from scratch; the Hebrew diacritiser and BEReshiT showed that adapting a modern model helps, but mainly within the genre of the adaptation data. Contamination is a recurring worry, because famous texts such as the Bible are likely already present in large models' training data.

Large language models as assistants. Recent papers use LLMs as correctors or extractors inside a pipeline rather than as end-to-end solvers, and keep humans in the loop. The AI Corpus Linguist let scholars query the Czech National Corpus in plain language, and its authors report that unannounced changes to the underlying hosted model made the tool unpredictable.

Resources

Many papers here release resources: annotated corpora of Syriac and Classical Armenian, the Jerusalem Corpus of Emergent Modern Hebrew, the Old German Reference Corpus, the BEReshiT and pre-segmented Rabbinic Hebrew models, a dataset of the dalet clitic's two functions, an Armenian inscription encoder for the EpiDoc standard, a vocabulary of Armenian inscription types, a layered annotation platform for Phoenician inscriptions, and GlobLingDiv, a country-level dataset relating linguistic diversity to digital support. Widely used external resources include the ETCBC database of the Hebrew Bible, Sefaria, the Perseus Digital Library, the Universal Dependencies Latin treebanks and the Royal Society Corpus.

Landmark papers in this library

Rule-Based Normalization of Historical Texts (2011). Derived weighted, context-sensitive spelling rules from two versions of the Luther Bible and checked candidates against a modern lexicon. It is the most cited work on the shelf and an early reference point for historical spelling normalisation.

Normalizing Early English Letters to Present-day English Spelling (2018). Compared five normalisation methods on letters from four centuries and showed that they fail on different words, so combining them has real potential, while grouping by period alone explains little.

MacBERTh (2021). Pre-trained BERT on English from 1450 to 1950 and evaluated it on tasks derived from the Oxford English Dictionary, with the largest gains on function words and good accuracy in dating quotations.

Lemmatization and POS-tagging for Classical Armenian, Old Georgian and Syriac (2020). Applied a joint neural lemmatiser and tagger to three under-resourced Christian Oriental languages and traced many remaining errors to forms with attached affixes.

Word Sense Induction with Attentive Context Clustering (2021). Combined a small word-embedding model with attention-weighted context vectors and density clustering, competitive on a standard English benchmark and able to split senses of Biblical Hebrew words.

New Technologies for Old Germanic (2014). Surveyed a verse-aligned corpus of Germanic Bibles, a reference corpus of Old High German and Old Saxon, and experiments projecting syntax from related languages.

Handling Stress in Finite-State Morphological Analyzers (2022). Showed that control symbols in the lexicon let finite-state tools handle Greek accents and the stress shift of Hebrew narrative verb forms.

Modeling the Dalet Clitic in Historical Hebrew Texts (2026). Pre-trained a Rabbinic Hebrew model with prefixes as separate tokens and used its predictions to label the two functions of an Aramaic particle across centuries of writing, revealing regional and authorial preferences.

Computational Methods for the Analysis of Complementizer Variability (2024). Parsed a very large Hebrew corpus to measure which verbs prefer each of the two complementizers, then used the results to spot deliberate stylistic choices in the novels of Amos Oz.

Style Classification of Rabbinic Literature (2022). Paired a genre classifier with a text-reuse filter to find candidate fragments of a lost midrash, a model of how computation can generate philological hypotheses for experts to test.

People

Avi Shmidman of Bar-Ilan University is the most prolific author on the shelf, with work on Emergent Modern Hebrew diacritisation, complementizer variation and the dalet clitic; Aynat Rubinstein of the Hebrew University of Jerusalem co-authored the first two, and Cheyn Shmuel Shmidman the third.

Stefanie Dipper of Ruhr University Bochum links the 2011 normalisation study, written with Marcel Bollmann and Florian Petran, to recent work on metaphor in German sermons. Carmen Schacht, also at Bochum, studies how architects and laypeople describe buildings differently.

Shai Gordin of Ariel University worked on Hebrew word senses with Amos Azaria and on annotation of Semitic inscriptions. Shmuel Liebeskind of Bar-Ilan and Chaya Liebeskind of the Jerusalem College of Technology built the Aramaic–Hebrew lexicon, and Shmuel Liebeskind later worked on responsa citations. Nachum Dershowitz of Tel Aviv University contributed the Genizah language segmentation and the search for lost Tanhuma material.

Hamest Tamrazyan and Emanuela Boros, associated with the École Polytechnique Fédérale de Lausanne, lead the Armenian epigraphy work.

The Helsinki group of Mika Hämäläinen, Jack Rueter, Jörg Tiedemann and Eetu Mäkelä wrote the Early English normalisation study. Christian Chiarcos and Maria Sukhareva led the older Germanic project. Daniel G. Swanson and Francis M. Tyers built the finite-state analysers.

Others include Sara Stymne of Uppsala on Latin parsing, Tim Van de Cruys of KU Leuven on Latin entities, Masayuki Asahara on Classical Japanese metaphor, Tom Lippincott on Torah sources, Alexander Fraser on Greek–Latin alignment, Frédérique Rey and Iglika Nikolova-Stoupak on BEReshiT, and Janis Pagel on LLMs and literary theory.

Institutions and places

Israel is the clear centre of gravity, through Bar-Ilan University, the Hebrew University of Jerusalem, Tel Aviv University, Ariel University and the Holon Institute of Technology, reflecting a strong tradition of computational work on Hebrew and Jewish texts. Germany follows, with Ruhr University Bochum and Humboldt-Universität zu Berlin. Other homes include EPFL, Ghent University, Indiana University and the French CNRS.

The venues are specialised. Most papers appeared at the SIGHUM workshop series on cultural heritage, social sciences and literature, and at LT4HALA; others come from the conferences and workshops on NLP for the digital humanities and from the LaTeCH workshop series.

Open questions

• Data scarcity. Gold annotation for ancient languages is small and expensive; weak supervision, transfer and synthetic data are partial answers.
• Genre and period shift. Models trained on one genre, such as prose, falter on another, such as poetry or ephemera, and period labels are crude proxies for variation.
• Contamination. Canonical texts are likely to appear in the training data of large models, which makes fair evaluation hard.
• Single-reference evaluation. Treebank conventions disagree, so a system can be penalised for defensible analyses.
• Interpretability for scholars. Humanists need to know why a tool made a decision; rule tables, readable weights and evidence spans remain valued for that reason.
• Reliance on hosted models. Tools built on commercial LLMs can change behaviour without notice, undermining reproducible scholarship.

Neighbouring topics

Nearly every paper here also stands on Language and Linguistics, and about half on Artificial Intelligence. A few papers on historical legal texts are shared with the Law shelf, and Greek–Latin alignment touches Machine translation.

By topic, the shelf is dominated by Hebrew & Aramaic and the wider Semitic languages, followed by Latin and Ancient Greek. Methodological links run to Language models & pretraining, Evaluation & benchmarks, Low-resource languages, Idioms & multiword expressions and Interpretability. Readers interested in the oldest written languages may also turn to Akkadian and cuneiform NLP.

Key works in this library

• Rule-Based Normalization of Historical Texts — spelling rules from aligned Luther Bibles.
• Normalizing Early English Letters to Present-day English Spelling — five normalisers compared.
• MacBERTh: Development and Evaluation of a Historically Pre-trained Language Model for English (1450-1950) — BERT for historical English.
• Lemmatization and POS-tagging process by using joint learning approach. Experimental results on Classical Armenian, Old Georgian, and Syriac — tools for three Oriental languages.
• Word Sense Induction with Attentive Context Clustering — senses of Biblical Hebrew words.
• New Technologies for Old Germanic. Resources and Research on Parallel Bibles in Older Continental Western Germanic — parallel Bibles and Old High German.
• Handling Stress in Finite-State Morphological Analyzers for Ancient Greek and Ancient Hebrew — accent-aware morphology.
• Language Classification and Segmentation of Noisy Documents in Hebrew Scripts — mixed-language Genizah texts.
• Automatic Construction of Aramaic-Hebrew Translation Lexicon — a lexicon from the Targums.
• A Thesaurus for Biblical Hebrew — semantic, phonetic and distributional links.
• Graph convolutional networks for exploring authorship hypotheses — sources of the Torah.
• Style Classification of Rabbinic Literature for Detection of Lost Midrash Tanhuma Material — finding lost midrash.
• NLP in the DH pipeline: Transfer-learning to a Chronolect — diacritising early modern Hebrew.
• Computational Methods for the Analysis of Complementizer Variability in Language and Literature: The Case of Hebrew "she-" and "ki" — complementizer choice in Hebrew.
• Modeling the Dalet Clitic in Historical Hebrew Texts: A New Prefix-Segmented BERT Model and Stylistic Analysis — an Aramaic particle in Rabbinic Hebrew.
• BEReshiT: an Ancient Hebrew Model based on DictaBERT — a model for Biblical Hebrew and the Dead Sea Scrolls.
• Language Models for the Restoration of Latin Legal Manuscripts — filling gaps in notarial deeds.
• THIVLVC: Retrieval Augmented Dependency Parsing for Latin — LLM correction of Latin parses.
• Contextual Probing for Low-Resource Named Entity Recognition in Latin — training-free Latin entities.
• Weakly Supervised Named Entity Recognition for Historical Texts — entity lists instead of full annotation.
• Evaluating Latin and Ancient Greek Sentence Alignment through Parallel Sentence Mining — Greek–Latin parallels.
• Degree Zero of Translation: Using Interlinear Baselines to Quantify Translator Intervention — measuring translator freedom.
• Armenian AutoEpiDoc: Automated Extraction and Encoding of Armenian Inscriptions into EpiDoc TEI/XML — digital Armenian epigraphy.
• Modeling Changing Scientific Concepts with Complex Networks: A Case Study on the Chemical Revolution — oxygen versus phlogiston.
• AI Corpus Linguist: More than a Year of Experience — plain-language corpus queries.