Etruscan: the readable language that cannot be fully translated

The corpus is an alphabetic one. The Etruscans took the Euboean Greek alphabet from the colonies at Pithekoussai and Cumae around 700 BC, kept 26 letters in their model abecedaria and used about 20 of them in practice, dropping the letters for voiced stops and the vowel o because the language lacked those sounds⁠[1]. Most inscriptions run right to left; word division by dots is common in later texts⁠[6]. The reference edition is Etruskische Texte (ET), published by Helmut Rix and his team in 1991 and enlarged in 2014 by Gerhard Meiser, Valentina Belfiore and Sindy Kluge with about 1,100 texts found since; each inscription carries a siglum built from the city and a type number, such as Ta 1.17 for a Tarquinia epitaph or LL for the linen book⁠[2]⁠[3]⁠[4]. The Corpus Inscriptionum Etruscarum, begun in 1893 and still incomplete, gives drawings and fuller apparatus by find-site⁠[23]. Every attested word form is indexed in the Thesaurus Linguae Etruscae, revised in 2009⁠[24].

Four long texts stand apart from the epitaphs. The Liber Linteus Zagrabiensis, a linen book torn into strips to wrap a mummy, holds about 1,330 legible words and 230 lines in twelve columns, perhaps 60 per cent of the original⁠[8]⁠[11]. The Tabula Capuana, a terracotta tile of about 470 BC, has about 390 legible words in ten sections, written boustrophedon, that is in alternating directions line by line⁠[9]. The Tabula Cortonensis, a bronze tablet of the second century BC, has about 200 words in 40 lines on two faces⁠[10]. The Cippus Perusinus, a travertine boundary stone of the third or second century BC, has 46 lines on two faces⁠[6]. To these belong the Pyrgi gold tablets, the only text paired with another language, and the bronze liver of Piacenza, whose 40 or so divine names in 16 outer regions anchor the Etruscan pantheon⁠[13]⁠[25].

The grammar is known in outline and much of it securely. Nouns inflect for case and number. There is a genitive in -s or -l, a locative in -i, a pertinentive in -si or -le used for recipients, and two layers of endings can stack, so that a word for "of the son of X" can be built by adding one genitive to another⁠[1]⁠[6]. Verbs show a past tense in -ce, a passive past in -che, imperatives, participles and a jussive, and little else; person and number are not marked on the verb⁠[1]. The numerals one to six are written on the Tuscania dice and are read as thu, zal, ci, sa, mach and huth, though which of sa and huth is four and which six is still argued⁠[6]. Personal names follow a fixed order of praenomen, family name and patronymic, which is why most epitaphs can be read in full⁠[15].

The lexicon is where knowledge stops. About 250 words have meanings that specialists treat as secure: kinship terms such as clan "son" and sech "daughter", avil "year", lupu "died", turuce "gave", the names of gods and of containers, a set of magistracies and ritual terms⁠[15]. Several hundred more have a guessed sense from context. These meanings come from three kinds of evidence: the thirty or so short Etruscan-Latin bilinguals, which fix kinship words and titles; glosses in Greek and Latin authors, who report for instance that aisar meant "gods"; and above all the combinatory method, which compares a word's position, its company and its endings across many inscriptions⁠[15]⁠[6]. The Piacenza liver and the ritual texts confirm one another on divine names, and the Pyrgi Phoenician text confirms the dedicator's name, the goddess Uni as Astarte and the word for three⁠[25]⁠[13]⁠[14].

Two methods have competed since the eighteenth century. The etymological method reads Etruscan words through a known language. Hebrew, Greek, Latin, Celtic, Basque, Armenian and Hittite were tried in turn; Zacharie Mayani read it as Albanian in 1961 and Mario Alinei as an archaic form of Hungarian in 2003⁠[15]⁠[6]. Each produces plausible readings for a few words and then fails on the morphology: Etruscan case endings, verb forms and numerals do not behave as the chosen language requires, and the readings do not carry over to the next inscription. The combinatory method, associated with Massimo Pallottino and the Italian school, abandons outside comparison and infers meaning from internal patterns, bilinguals and glosses⁠[26]⁠[15]. It is slow, and it is what has produced every accepted result, including the numerals, the kinship terms and the structure of the ritual calendars⁠[6].

The long texts have each had full-length commentaries. For the linen book, van der Meer in 2007 and Belfiore in 2010 agree that it is a liturgical calendar prescribing offerings to named gods on dated days, written by one hand in northern Etruria between about 250 and 150 BC; they agree on two month names, acale for June and celi for September, and on little else at the level of the sentence⁠[11]⁠[8]⁠[12]. Cristofani read the Capua tile as a festival calendar of ten months⁠[9]. Agostiniani and Nicosia read the Cortona tablet as a record of a land transaction between the Cusu family and Petru Scevas, with lists of witnesses, and Wylin used the shorter Pyrgi tablet to recover its drafting formulae⁠[10]⁠[27]⁠[28]. Translations of these texts that go beyond genre, names and formulae, including several published full versions of the linen book, are not accepted by Etruscologists because they cannot be checked against anything.

Computational work is recent and thin. A 2023 survey of machine learning for ancient languages found no open Etruscan corpus and no modelling⁠[29]. The same year Vico and Spanakis released the Larth dataset, 7,139 inscriptions with 2,891 English translations drawn from the Etruscan Texts Project and CIE I, and trained translation models that reached a BLEU score of 10.1, which is far below usable quality⁠[5]⁠[30]. The OpenEtruscan corpus of 2026 merges and cleans these sources to 6,567 rows and adds geographic links, but its text classifiers are scored against machine-generated labels rather than expert ones⁠[31].

On affiliation, four positions are live. The default among specialists is that Etruscan belongs to a small Tyrsenian family with Raetic and Lemnian and has no demonstrable relatives beyond them. Rix proposed the grouping in 1998 and Schumacher's Raetic corpus supplies the correspondences: the genitive -s, the second genitive -a, the past participle -ce beside Raetic -ku, zal for two in both, and shared name-building suffixes⁠[16]⁠[17]. The Lemnos stele of about 550 BC and the Hephaistia inscription of 2009 share with Etruscan the dative endings -si and -ale, a past tense in -ai and the formula for stating a person's age⁠[18]⁠[32]. The second position, argued by Beekes, is that the family is an offshoot of a pre-Greek Aegean and western Anatolian population that moved to Italy around 1200 BC, as Herodotus's Lydian story preserves in distorted form⁠[21]⁠[19]. The third, Dionysius's, is that the Etruscans and their language are native to Italy and the Lemnians are Etruscan settlers who went east; the 2021 genomes, which show no Anatolian admixture in Iron Age Etruria, favour this for the people without settling it for the language⁠[20]⁠[22]. The fourth is the family of etymological readings, Albanian, Hungarian and the rest, which have no standing among epigraphers⁠[15].

On the long texts, the accepted reading is one of genre and structure. The linen book and the Capua tile are ritual calendars; the Cortona tablet and the Perugia stone are legal records concerning land, with family names, boundaries and witnesses⁠[11]⁠[9]⁠[10]⁠[6]. Within that frame individual words are read with varying confidence, from the secure month names and divine names down to verbs of offering whose exact sense is argued, and to the many words that occur once and cannot be placed⁠[8]. No reading of any of the four texts is accepted sentence by sentence.

A true bilingual would settle most of it. A Latin or Greek translation of a connected Etruscan text, even a dozen lines, would fix verbs and particles that the epitaphs never show, and the formulae of the four long texts are repetitive enough that a few fixed sentences would cascade⁠[15]⁠[6]. Etruscan books existed, Roman writers cite them, and linen books were buried with their owners, so a second linen book from an Egyptian mummy or a waterlogged tomb is possible rather than fanciful⁠[11]. Short of that, more Raetic and Lemnian is the next best thing. Each new Tyrsenian text tests the shared morphology, and a long Lemnian or Raetic text with a Greek or Latin parallel would work almost as well as an Etruscan one⁠[32]⁠[17].

Three tests can be run now. The first is to put the combinatory method on a quantitative footing: for every word in ET, compute its distribution over text types, positions and collocates and see which guessed meanings predict new attestations and which do not, using the open Larth and OpenEtruscan files as a start and ET² as the check⁠[5]⁠[31]. The second is to test the Tyrsenian correspondences statistically against a null of chance resemblance, as has been done for other proposed families; the Raetic and Lemnian corpora are small but the morphological matches are specific. The third concerns origins rather than language: ancient DNA from Lemnos and from the Raetic Alps, compared with the Etruscan transect, would show whether the three Tyrsenian populations share recent ancestry or only a language⁠[22]. None of these will translate the linen book, but they will say what kind of evidence could.