1 Overview
Repository location
Context
Inner-biblical parallels, the passages in which one biblical author draws on, reworks, or echoes another, have been catalogued by biblical scholars since antiquity. The biblical book of Chronicles and its parallels in Samuel and Kings have been the subject of extensive synoptic study (Bendavid, 2010; Endres et al., 1998), while the analysis of poetic parallelism within and across the Psalms, prophecy, and wisdom literature has produced its own substantial body of scholarship (Berlin, 2008; Fokkelman, 2001; Kugel, 1981; Tsumura, 2023; Watson, 1994). What has not existed is a machine-readable resource that pairs the Hebrew text of these previously identified parallels with morphological annotation suitable for computational research.
T’OMIM (Tanakh Observable Matches of Intertextual Mimesis; from Hebrew תאומים, “twins”)1 provides such a resource. It bridges two scholarly traditions, narrative-synoptic comparison and poetic parallelism studies, by aligning the catalogued pairs of both traditions to the ETCBC’s BHSA (van Peursen et al., 2015), a fully parsed digital edition of the Hebrew Bible. The dataset was constructed in the context of the author’s research on computational detection of biblical parallels using finetuned sentence-transformer models (Smiley, 2026), but is released independently of any model so that other researchers can rebuild without re-curation.
2 Method
T’OMIM is built from two source streams: published catalogs of parallel passages and the ETCBC BHSA morphological database. The narrative subset draws its pair identifications from Bendavid’s (2010) compendium of biblical parallels and from Endres et al.’s (1998) tabulation of Chronicles and its synoptic parallels in Samuel, Kings, and related texts. The poetic subset draws on Berlin (2008), Fokkelman (2001), Kugel (1981), Watson (1994), and Tsumura (2023), each of which catalogues paired half-verses (cola) that exhibit grammatical, lexical, or semantic parallelism. In both cases the underlying Hebrew text and morphological tags come from ETCBC.
Steps
Construction proceeded along two pipelines that converge on a shared word-level schema. For the narrative pairs, the Chronicles synoptic catalog was joined to ETCBC verses by canonical reference (book, chapter, verse), yielding 554 verse-level narrative pairs. The poetic pipeline required a different approach. Because scholarly half-verse citations often refer to spans below the granularity of the BHSA half-verse labels (for example, the third colon of a tricolon, or a sub-half-verse phrase), pair identification was performed by character-level alignment: scholarly references were matched against ETCBC word data after stripping niqqud, cantillation marks, and the maqef. This procedure recovers c- and d-cola and permits sub-half-verse matches that label-only joins miss. This process yielded 256 verse-level poetic pairs.
Each verse or half-verse was then expanded to its constituent ETCBC tokens. From the raw BHSA fields, 18 columns were dropped where more than 90% of the values were null at word granularity (most of these encode aggregate properties of clauses or phrases), leaving the 58 columns documented in the dataset description on the Zenodo record. The expansion produces 25572 word-level rows for the narrative subset and 2437 for the poetic subset.
Sampling strategy
No sampling was performed. Every parallel identified in the source catalogs was included, subject to successful alignment to ETCBC. The dataset is therefore exhaustive within its sources, not a stratified or random sample. This design supports its use as a coverage benchmark for parallel-passage retrieval, but it also means that the genre and book distributions of T’OMIM follow the distributions of the underlying scholarly traditions (see Section 4 for limitations).
Quality control
Thirty-two structural checks were defined to verify the integrity of the build. These cover row counts against expected pair totals, schema consistency between narrative and poetic tables, the integrity of pair_id joins between the verse-level and word-level tables, ETCBC node-ID continuity within reconstructed verses, and the absence of orphaned tokens or broken half-verse references. All checks pass for the released version. In addition, a stratified sample of 50 pairs was manually inspected against the published source catalogs to confirm that the stored Hebrew text matches each catalog reference.
3 Dataset Description
Repository name
Zenodo
Object name
T’OMIM: Tanakh Observable Matches of Intertextual Mimesis (v2).
Format names and versions
Apache Parquet (canonical, written with PyArrow), CSV (UTF-8), and Excel (XLSX). All four tables are provided in all three formats.
Creation dates
2024-08-01 to 2026-03-22.
Dataset creators
David M. Smiley (sole curator). Upstream credit for the Hebrew text and morphological annotation is owed to the Eep Talstra Centre for Bible and Computer (ETCBC), whose BHSA database is the source of all linguistic data (van Peursen et al., 2015).
Language
Classical Hebrew. Column names, English glosses, and metadata fields are in English.
License
Creative Commons Attribution 4.0 International (CC-BY-4.0).
Publication date
2026-03-22.
The dataset comprises four Parquet files: narrative_pairs_verse.parquet (554 rows), narrative_pairs_word.parquet (25572 rows), poetic_pairs_verse.parquet (256 rows), and poetic_pairs_word.parquet (2437 rows). The verse-level and word-level tables are linked by a sequential pair_id and a side field (“source” or “target”). Table 1 illustrates this two-level structure; the full 58-column word-level schema is documented in the dataset description on the Zenodo record. All four tables (narrative_pairs_verse, narrative_pairs_word, poetic_pairs_verse, and poetic_pairs_word) are available on Zenodo in all three formats: Apache Parquet (the canonical version), CSV, and Excel. The CSV and Excel copies have been verified cell for cell against the Parquet originals, including all pointed Hebrew text.
Table 1
Example pairs from T’OMIM. Panel A shows one narrative and one poetic pair at verse granularity. Panel B expands the poetic pair to its word-level tokens, displaying 12 of the 58 columns in the word-level schema.
| Panel A: Verse-level rows (from narrative_pairs_verse and poetic_pairs_verse) | |||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| TYPE | PAIR_ID | SOURCE_REF | TARGET_REF | SOURCE_TEXT | TARGET_TEXT | REFERENCE | |||||
| Narr. | 10 | 1 Sam 31:5 | 1 Chr 10:5 | וַירְַּא נֹשֵׂא־כֵלָיו כִּי מֵת שָׁאולּ וַיפִֹּּל גםַּ־הואּ עַל־חַרְבֹּו וַימָָּת עִמֹּו | וַיַּרְא נֹשֵׂא־כֵלָיו כִּי מֵת שָׁאוּל וַיִּפֹּל גַּם־הוּא עַל־הַחֶרֶב וַיָּמ ̵ֹת | Endres and Ben-David | |||||
| Poet. | 150 | Isa 1:10a | Isa 1:10b | שִׁמְעוּ דְבַר־יְהוָה קְצִינֵי סְדֹם | הַאֲזיִנו תֹּּורַת אֱלֹהֵינו עַּם עֲמֹרָה | Kugel, 31 | |||||
| Panel B: Word-level excerpt (from poetic_pairs_word; poetic pair, Isa 1:10a // Isa 1:10b) | |||||||||||
| PAIR_ID | SIDE | G_WORD_UTF8 | LEX_UTF8 | GLOSS | SP | VS | VT | PS | GN | NU | ST |
| 150 | source | שִׁמְעוּ | שׁמע | hear | verb | qal | impv | p2 | m | pl | — |
| 150 | source | דְבַר | דבר | word | subs | — | — | — | m | sg | c |
| 150 | source | יְהוָה | יהוה | YHWH | nmpr | — | — | — | m | sg | a |
| 150 | source | קְצִינֵי | קצין | chief | subs | — | — | — | m | pl | c |
| 150 | source | סְדֹם | סדם | Sodom | nmpr | — | — | — | — | sg | a |
| 150 | target | הַאֲזִינוּ | אזן | listen | verb | hif | impv | p2 | m | pl | — |
| 150 | target | תֹּורַת | תורה | instruction | subs | — | — | — | f | sg | c |
| 150 | target | אֱלֹהֵינוּ | אלהים | god(s) | subs | — | — | — | m | pl | a |
| 150 | target | עַם | עם | people | subs | — | — | — | m | sg | c |
| 150 | target | עֲמֹרָה | עמרה | Gomorrah | nmpr | — | — | — | — | sg | a |
4 Reuse Potential
T’OMIM is designed to support research at the intersection of computational linguistics, biblical studies, Hebrew literary studies, and digital humanities. Three audiences are envisioned.
For computational researchers, the dataset provides labelled positive pairs for training and evaluating models of semantic similarity, paraphrase detection, and text-reuse retrieval in Biblical Hebrew, a low-resource language for which annotated benchmarks remain scarce. The morphological layer enables ablation studies that isolate the contribution of inflectional information from raw lexical overlap. T’OMIM has already been used in this capacity to finetune a new Classical Hebrew language model, called MiqraBERT (Smiley, 2026), for inner-biblical parallel detection.
For linguists, the word-level tables expose the morphological behavior of paired passages. Comparative work on how parallelism is realized at the level of verbal stem (binyan), tense, person, gender, and number becomes tractable at scale. Such comparisons have traditionally been pursued through close reading; however, T’OMIM permits distant readings over the full set of catalogued pairs.
For biblical scholars, T’OMIM consolidates two scholarly conversations (narrative synoptic comparison and poetic parallelism) into a single machine-readable index of known parallels. The dataset supports concordance-style queries (e.g., “which reformulations in Chronicles preserve the verb but change the binyan?”). It also supports exegetical work that would otherwise require laborious manual alignment of source and target passages and provides a starting index for reception-history studies.
Four limitations bear directly on reuse. First, T’OMIM is catalog-bound: it contains only previously identified parallels and is therefore appropriate as a benchmark for parallel-detection systems, but it cannot be used to validate the discovery of novel parallels not present in its source catalogs. Second, the genre distribution is uneven: narrative pairs are heavily Chronicles-centred, reflecting the focus of the underlying synoptic tradition, and poetic pairs overrepresent the Psalms and the major prophets. Researchers seeking balanced coverage of, for example, narrative material outside Chronicles or wisdom poetry should expect underrepresentation. Third, no inter-annotator agreement metric is reported, because each pair derives from a single published scholarly source rather than from fresh annotation. Users requiring a confidence measure should treat the source citation in the “reference” column as the relevant provenance signal. Fourth, T’OMIM contains only positive pairs. Any discriminative training task requires the user to construct negative pairs.
Note
AI Declaration
Claude Code was consulted in the process of editing the manuscript draft for spelling mistakes, inconsistencies, and grammatical structure/errors, as well as LaTeX conversion and troubleshooting.
Acknowledgements
The author thanks the Eep Talstra Centre for Bible and Computer (ETCBC) for making the BHSA dataset openly available under a permissive license, without which T’OMIM could not have been assembled.
Author Contributions
Conceptualization, Methodology, Software, Data curation, Validation, Writing — original draft, Writing — review and editing: David M. Smiley.