Daejeon, July 27–31
The Bible has been translated more often than any other text, both across languages and, within a number of them, across centuries of retranslation. Due to this availability and its standardized verse structure, it served as a foundational dataset for early Statistical Machine Translation (SMT), where researchers relied on the assumption of semantic equivalence between aligned verses (Mueller et al. 2020). However, the landscape of NLP has evolved. In the era of Large Language Models trained on terabytes of text, the Bible contributes only marginally to pre-training mixes, and its narrow thematic and stylistic register is unrepresentative of general language. This shift in role is itself an opportunity: the same scale that makes the Bible obsolete as training data makes it newly tractable as an object of study, with AI systems serving as the analytical instrument rather than the consumer of the text. With its deep history, the Bible is well suited for investigating phenomena ranging from diachronic language change to synchronic translation variation.
Existing biblical corpora were not designed for this analytical depth. Most prior efforts focused on linguistic breadth, aiming to cover as many languages as possible to support typological research. Consequently, the coverage of translation history within individual languages is rarely a primary design goal; when multiple versions are included, their selection is often incidental rather than representative. Furthermore, such corpora rarely annotate canonical identity at the edition or revision level, so closely related versions are either silently merged or filtered out, erasing the fine-grained variation of translation families.
To address this gap, we introduce targum, a corpus that prioritizes vertical depth over linguistic breadth. We aggregated 651 New Testament translations from 13 sources across five languages: English, French, Italian, Polish, and Spanish. Instead of removing duplicates, we retained them and manually annotated every text to map it to a specific canonical edition and revision year. This design shifts the definition of uniqueness to the researcher: using our metadata, a user can decide whether to treat the King James Version of 1611 and the 1769 revision as duplicates to be filtered, or as distinct historical data points to be analyzed.
The name pays tribute to the ancient Aramaic Targums of the Hebrew Bible (תרגום), which combined literal rendering with dynamic expansion for the target audience (Alexander 1992); our corpus likewise spans translations ranging from formal literalism to dynamic modern paraphrase.
Biblical corpora generally follow two paradigms. The first maximizes linguistic breadth to support typological studies and low-resource NLP. Prominent examples include the collections by Christodouloupoulos / Steedman (2015) and Mayer / Cysouw (2014), as well as the Johns Hopkins University Bible Corpus (McCarthy et al. 2020) and the eBible Corpus (Akerman et al. 2023). While these resources provide essential coverage for thousands of languages, their primary objective is cross-lingual diversity. Consequently, they do not aim to systematically capture the full translation history within individual languages; even when multiple versions are present, they are not selected to represent a continuous historical or stylistic spectrum. We provide a detailed quantitative comparison of this coverage in Section 4.
The second paradigm focuses on vertical depth to enable diachronic or stylistic analysis. The EDGeS Diachronic Bible Corpus, for example, aggregates 36 translations across four Germanic languages to study verb constructions (Bouma et al. 2020). Similarly, Carlson et al. (2018) compiled 34 English versions to evaluate prose style transfer. targum extends this depth-focused approach. Rather than targeting a specific linguistic feature, we provide a general-purpose resource for comparative analysis across five major European translation traditions.
We aggregated translations from 13 distinct online sources. Table 1 details the complete list of origin sites and the specific volume of texts acquired from each for every target language. The acquisition pipeline operated in three stages. First, an indexer cataloged available translations to create a timestamped snapshot. Second, a scraper downloaded the raw files (HTML/JSON) for the 260 New Testament chapters. Third, a parser extracted verse-level text, handling variable segmentation schemes (e.g., ranges like 1–2, 3–6c) to output structured JSON and TSV files.
| Source | eng | fra | ita | pol | spa | Total |
| bible.audio | 3 | 15 | 0 | 0 | 1 | 19 |
| bible.com | 75 | 19 | 7 | 8 | 28 | 137 |
| bible.is | 18 | 8 | 1 | 2 | 15 | 44 |
| biblegateway.com | 64 | 4 | 5 | 3 | 19 | 95 |
| biblehub.com | 46 | 0 | 0 | 0 | 0 | 46 |
| bibles.org | 72 | 6 | 3 | 3 | 14 | 98 |
| biblestudytools.com | 37 | 0 | 0 | 0 | 0 | 37 |
| bibliepolskie.pl | 0 | 0 | 0 | 20 | 0 | 20 |
| crossbible.com | 12 | 1 | 0 | 0 | 1 | 14 |
| ebible.org | 24 | 4 | 2 | 2 | 5 | 37 |
| jw.org | 1 | 1 | 1 | 1 | 1 | 5 |
| laparola.net | 9 | 3 | 11 | 0 | 2 | 25 |
| obohu.cz | 29 | 17 | 3 | 9 | 16 | 74 |
| Total | 390 | 78 | 33 | 48 | 102 | 651 |
Table 1: Number of translations sourced from each site for each language in the targum corpus.
Raw scraped data frequently contained inconsistent naming or lacked explicit versioning information. To resolve this, we manually mapped each of the 651 scraped instances to a canonical entry, assigning a canonical_id (identifying the distinct work, e.g., “King James Version”) and a canonical_revision_year (identifying the specific edition, e.g., “1611”).
Determining the exact revision often required extensive verification. When source websites provided ambiguous or incomplete metadata, we cross-referenced the text with external resources, including publisher websites, Wikipedia, and digital libraries. For older historical translations where metadata was scarce, we validated the edition by comparing the scraped text against scanned facsimiles of the original printed books. We attempted to automate this metadata extraction using LLMs, but models frequently hallucinated publication dates or conflated distinct revisions. Consequently, we relied entirely on manual verification to ensure that users can reliably distinguish between a new translation and a minor reissue.
The pre-computed lexical similarity scores released with the corpus support a quantitative check on this scheme. Pairs sharing both canonical_id and canonical_revision_year cluster tightly near one (median 0.996), pairs sharing only canonical_id form a wider but still compact distribution (median 0.99), and unrelated within-language pairs centre near 0.53; the three tiers separate cleanly.
The redundancy of our collection supported quality control. By calculating the edit distance between duplicates sharing a canonical_id, we identified discrepancies. Most often, these highlighted parsing irregularities, allowing us to refine our site-specific extraction algorithms. In a handful of cases, however, this comparison revealed issues within the source texts themselves, such as typos or missing verses. Such anomalies would not surface from a single source.
The complete corpus is publicly available at https://github.com/mrapacz/targum-corpus. Of the 651 translations, 237 are in the public domain and 65 carry open licenses, distributed as raw text with metadata. The remaining 349 are copyrighted and released only as derived artefacts — verse- and chapter-level Qwen3-Embedding (Zhang et al. 2025) vectors (0.6B and 8B) and pairwise similarity scores — with text available upon request for non-commercial research; we will remove any translation upon a verified copyright-holder request.
Table 2 compares targum against existing resources. The corpus increases the number of available unique translations across all target languages. For English, we provide 194 unique versions, a 5.0× increase over the 39 versions in Mayer / Cysouw (2014). Polish coverage shows comparable expansion, growing from 6 to 29 unique versions (4.8×). The Romance languages grow at similar rates: Italian increases from 7 to 17 (2.4×), Spanish from 22 to 53 (2.4×), and French from 17 to 41 (2.4×).
| Corpus | eng | fra | ita | pol | spa |
| Christodouloupoulos / Steedman (2015) | 2 | 1 | 1 | 1 | 1 |
| Mayer / Cysouw (2014) | 39 | 17 | 7 | 6 | 22 |
| Akerman et al. (2023) | 24 | 4 | 2 | 2 | 5 |
| Targum (Total) | 390 | 78 | 33 | 48 | 102 |
| Targum (Unique) | 194 | 41 | 17 | 29 | 53 |
Table 2: Number of translations per language in targum compared to major prior corpora.
This depth results from our hybrid sourcing strategy. While global aggregators like bible.com provided the bulk of English (75) and Spanish (28) texts, national repositories were critical for other languages. For example, bibliepolskie.pl provided 20 Polish translations not found on international sites, and laparola.net contributed 11 Italian versions.
We analyzed the temporal distribution of the unique translations. Quantitatively, the dataset exhibits a distinct “digitization bias,” with the proportion of unique editions published in the last 50 years ranging from 47% (Italian) to 83% (Spanish). Despite this modern concentration, the corpus captures key historical phases, with the oldest English text dating to 1525. The distribution also quantifies historical discontinuities: Polish and Spanish each exhibit a gap of roughly 250 years between their early 17th-century canonical versions and the mid-19th-century revival.
To quantify internal diversity, we computed pairwise similarity for every translation pair within each language, aggregating chapter-level scores across all 27 New Testament books. Semantic similarity is the cosine distance on chapter-level Qwen3-Embedding-8B (Zhang et al. 2025) vectors; the corpus also distributes 0.6B vectors and verse-level versions of both for users with different compute or granularity requirements. Lexical similarity is computed as 1 − LD(x₁, x₂) / max(|x₁|, |x₂|), where LD is the Levenshtein distance.
Within-language semantic similarity is high (mean 0.955), reflecting the shared narrative core of the corpus, but the per-language averages already hint at structural differences: English sits at the low end (0.943) because its larger and more stylistically diverse collection includes paraphrastic outliers, while French and Italian — with smaller, more homogeneous collections — sit at the top (0.965). Cross-language semantic similarity is lower (mean 0.823) and traces a recognisable typological pattern: among the Romance languages, Italian–Spanish is the closest pair (0.863), followed closely by English–French (0.862), while Polish — the only Slavic language in the corpus — is the most distant from every other language, with French–Polish the lowest pair at 0.766. Lexical similarity captures finer stylistic variation that the embeddings smooth over, and is most informative at the level of individual translation pairs rather than aggregate means.
targum shifts the focus of corpus construction from covering many languages to representing translation history within a select few. This approach supports research directions that sparse collections do not. Researchers can now trace how specific text lineages evolve over centuries, or compare new translations against the established canon to measure stylistic differences.
By combining these texts with manual annotations, targum allows us to treat the New Testament's translation history as a complex data object, spanning centuries, theological traditions, and linguistic styles.
targum corpus on GitHub: https://github.com/mrapacz/targum-corpus