Sphragis

Sphragis (σφραγίς/σφρηγίς, "sigil") is a benchmark for Ancient Greek (grc) prose and verse authorship attribution (AA), where the input features are human-annotated CoNLL-U trees, which include both semantics (lemma) and syntax (POS, morphology and dependency relation). For verse, the features pair the CoNLL-U with metrical annotation, such as syllable lengths and brevis in longo. The benchmark defines three tasks: 1, 10, and 100-sentence (= line for the verse-metre genre) chunk attribution. Sphragis is thus the first benchmark that enables comparison of all hitherto suggested models and methods of grc automatic AA, regardless of feature source, and further, because of the three chunk-length tasks, fair and informative comparison of both models trained on sparse feature vectors of rates or counts, and those trained on dense representations (e.g. from transformers).

For the archaic Greek poets, "sphragis" (or "sphregis" in some dialects) was a term of art which roughly equivalent to the modern practice of "authorial watermarking", referring to text composition that intentionally aims at making it harder for adversaries to falsely pass off the work as their own. The phenomenon is one of the earliest chapters in the history of AA. (See Theognidea, 19-30)

To enable the double feature set of the verse genre, the benchmark consists of the intersection of all open grc treebanks with the open corpus of scanned verse [Hypotactic)(https://hypotactic.com/).

Configurations

Configuration	Authors	Train unit	Train rows	Evaluation unit	Validation rows	Test rows
prose_1	12	sentence	25,730	sentence	2,800	2,800
prose_10	12	sentence	25,730	10-sentence chunk	280	280
prose_100	12	sentence	25,730	100-sentence chunk	28	28
verse_sentence_1	2	sentence	11,914	sentence	1,400	1,400
verse_sentence_10	2	sentence	11,914	10-sentence chunk	140	140
verse_sentence_100	2	sentence	11,914	100-sentence chunk	14	14
verse_metre_1	2	scanned line	21,818	scanned line	2,600	2,600
verse_metre_10	2	scanned line	21,818	10-line chunk	260	260
verse_metre_100	2	scanned line	21,818	100-line chunk	26	26

The suffix gives the evaluation chunk size. Within each genre, _100 defines one shared source corpus for all three tasks: an author is retained only when that author independently has at least 100 source rows in both validation and test. Training remains at the atomic sentence or line level so that training-time chunking remains an experimental choice. _1 publishes the shared evaluation rows atomically, while _10 and _100 regroup those exact same rows into fixed benchmark chunks. Thus author inventory and total source text are identical across task sizes within a genre.

Evaluation selection uses seed 776. For each retained author and split, the unavoidable n mod 100 source rows are selected for exclusion once by a stable hash. The shared retained rows are put into their best available textual order within each work for each chunking task. Complete single-work chunks are emitted first; only residual work tails are combined, and such rows are marked by chunk_is_mixed_work. Chunk IDs, boundaries and exclusions are therefore reproducible. The complete manifest is in metadata/dataset_variants.json.

The verse source pool is constructed through exact syntax/metre alignment. Task bottlenecking is then applied independently to verse_sentence and verse_metre, whose atomic units and splits differ. Evaluation chunks preserve all constituent links and provenance, but syntax and metre chunk boundaries are intentionally independent.

from datasets import load_dataset

prose = load_dataset("Urdatorn/sphragis", "prose_10")
verse_syntax = load_dataset("Urdatorn/sphragis", "verse_sentence_10")
verse_metre = load_dataset("Urdatorn/sphragis", "verse_metre_10")

Inclusion rule for verse

A contiguous verse stretch is included only if:

1. its beginning and end are both positions where a Hypotactic line boundary coincides with a human treebank sentence boundary;
2. the complete normalized text between those boundaries is identical in the two annotation streams; and
3. every intervening line has a metrical scan and every intervening sentence has a retained CoNLL-U tree.

The aligner computes a character-level edit path between the normalized streams and retains exact components between joint boundaries. A line may therefore span several sentences, and a sentence may span several lines. Text at the edge of a treebank selection is excluded until the next joint boundary; there are no partial or fuzzy alignments. This retains 3,729 lines crossing a treebank sentence boundary.

The resulting verse material covers the Iliad, Odyssey, Hesiod's Theogony and Works and Days, Aeschylus' Persians and Seven Against Thebes, and Semonides. Disputed passages 861–874 and 1005–1078 of Seven Against Thebes are excluded at complete alignment-component boundaries.

Splits

Splits use random seed 776 and begin with independent stratification within each author/work group: approximately 80% train, 10% validation and 10% test by row. Groups with 3–9 rows receive one validation and one test row; groups with fewer than three stay in training. Verse syntax and metre rows are split independently, so linked alignment components need not have identical split labels. The 100-row eligibility and remainder bottleneck is then applied to all three task sizes. Training is left atomic, while the equally sized validation and test source sets are exact multiples of 100 per author. This makes the published source-row ratio slightly different from the initial 80/10/10 split. A stricter work-held-out evaluation can be added as a later track.

Principal columns

All configurations contain:

• id, author, work, work_id, genre, text;
• conllu, cts_urn, passage, treebank_source;
• source_records: JSON containing every contributing upstream record, revision, URL, license, annotation provenance and syntax scheme;
• licenses, dedup_key, and split.

The _10 and _100 configurations additionally contain chunk_size, chunk_target_size, ordered constituent_ids, JSON constituent_provenance, chunk_work_ids, chunk_works, and chunk_is_mixed_work. Training rows have chunk_size = 1; validation and test rows always have chunk_size equal to the configuration suffix. Aggregated text and conllu follow constituent order, while constituent provenance retains the original passage-level locations.

verse_sentence additionally contains metrical_line_ids and a JSON list of the full overlapping metrical_lines; an overlap is marked full or partial. verse_metre contains the native list parent_sentence_ids, book, poem_sequence, line_number, raw Hypotactic metre, flattened syllable-level annotations, and the originating filename. Its conllu value is a valid multi-sentence CoNLL-U document when a line crosses a sentence boundary. Both verse configurations contain alignment_component_id and an index giving row order within that component.

genre has exactly three values in this release: prose, verse_sentence, and verse_metre.

Dependency representation

Existing UD data is retained verbatim. Native AGDT/Arethusa trees are serialized as CoNLL-U while preserving heads. Their dependency relations are conservatively mapped to the Universal Dependencies v2 inventory; opaque source-specific categories fall back to dep. The lossless original NativeRel and NativeHead values remain in MISC for audit and future mapping improvements. Repairs required after removing artificial ellipsis nodes—or to break an upstream beta self-loop/cycle—are explicitly marked HeadRepair=Yes.

XPOS is normalized across every source to the nine-position Ancient Greek AGDT/Perseus tag convention. Existing valid positional tags are retained; incompatible source-specific tags are reconstructed from UPOS and FEATS, with unavailable distinctions represented by -. UPOS and FEATS remain the canonical UD fields.

Every distinct published conllu document is loaded during testing with the official CoNLL 2018 shared-task load_conllu parser. This checks sentence termination, ten-column structure, token IDs, heads, roots and dependency cycles across all nine configurations.

To prevent authorship-label leakage, model-facing conllu values retain only the non-identifying # text = ... comment. Comments carrying sent_id, source, cts, passage, or newdoc id are removed. The same information remains available for audit in the dataset's dedicated provenance columns. Token-level MISC identifiers (Cite, Ref, and LId) are also removed; only NativeRel, NativeHead, HeadRepair, and SpaceAfter are retained there.

Deduplication

Raw prose input comprised 72,336 sentences. Authorship curation retained 55,273 source sentences before deduplication. Exact normalization-based deduplication then removed 13,510 repeated rows, and 299 rows in 53 exact-text groups were omitted because their retained author labels conflict, leaving 41,464 rows. Normalization lowercases, maps final sigma, removes diacritics and retains letters only. The chosen parse follows a deterministic source priority, while all exact-source provenance is merged into source_records.

This is aggressive exact deduplication, not passage-level fuzzy deduplication. Segmentation or edition differences can therefore leave near duplicates for future review.

Sources and licensing

The corpus draws on AGDT 2.1, UD Perseus, UD PROIEL, UD PTNK, Gorman's Greek Dependency Trees, Pedalion, Harrington Trees, and Hypotactic.

This is a mixed-license dataset, including NonCommercial and ShareAlike data. Consult LICENSES.md and the licenses recorded on every row. Source commits for this release are frozen in metadata/source_revisions.json.

Conservative authorship policy

The benchmark favors reliable training labels over coverage. It excludes anonymous and unknown authors, Pseudo-* and traditional attributions, fragment-labelled works, corporate or mediated authorship, and works or passages whose attribution is substantially disputed. This includes disputed material transmitted inside the Demosthenic, Antiphontic, Lysian, Platonic, Aeschylean, Isocratean and Xenophontic corpora, as well as the pseudo-Hesiodic Shield of Heracles. The core Theogony and Works and Days are retained under Hesiod. Romans is retained under Paul, while disputed and anonymous New Testament works are excluded.

The Iliad and Odyssey are retained as distinct conventional corpora under the author labels Homeric-Iliad and Homeric-Odyssey; the labels do not claim a shared biographical author. Every work labelled Fragments is excluded. Machine-readable reasons and pre-deduplication counts are recorded in metadata/build_report.json.

[1] bla [2]