Dragoman

--- language: - grc - el - en license: cc-by-4.0 library_name: transformers thumbnail: https://huggingface.co/open-greek/dragoman/resolve/main/thumbnail.png tags: - word-alignment - ancient-greek - modern-greek - simalign - xlm-roberta - homer - iliad - digital-humanities base_model: UGARIT/grc-alignment pipeline_tag: feature-extraction --- # Dragoman: Diachronic Greek Word Alignment Diachronic word alignment model for Ancient Greek (grc), Modern Greek (el), and English (en). Fine-tuned from [UGARIT/grc-alignment](https://huggingface.co/UGARIT/grc-alignment) on Iliad parallel text with contrastive alignment training. Dragoman extends UGARIT's AG-EN alignment to also handle AG-MG (Ancient to Modern Greek) alignment, with significant improvements on both axes. The model is designed for [Iliad Aligned](https://iliadaligned.com) but generalizes to any AG-EN or AG-MG parallel text. ## Model Details - **Base model**: [UGARIT/grc-alignment](https://huggingface.co/UGARIT/grc-alignment) (XLM-RoBERTa, 278M params) - **Fine-tuning**: Contrastive alignment loss (InfoNCE) on layer 8 embeddings - **Lemma head**: ~200K param disambiguation module trained alongside alignment - **Training data**: ~120K pairs from Iliad silver standard, Cunliffe + LSJ lexicons, Perseus prose, NT (OpenGNT), UGARIT gold, AG-MG gold, Attic drama (15.8K theatre pairs) - **Training time**: ~20 min on RTX 2080 Ti (training only; full pipeline including inference is ~2 hours) - **License**: CC-BY-4.0 (following UGARIT) ## Intended Use Word-level alignment between: - Ancient Greek ↔ English translations - Ancient Greek ↔ Modern Greek translations (including Katharevousa) Primary use case: parallel text readers, digital humanities tools, and Greek NLP pipelines. The model produces word-level alignment pairs via [simalign](https://github.com/cisnlp/simalign)'s argmax algorithm. ## Alignment Results ### AG-EN | Model | Precision | Recall | F1 | |-------|-----------|--------|------| | UGARIT/grc-alignment (base) | 0.586 | 0.723 | 0.648 | | **Dragoman + post-processing** | **0.708** | **0.908** | **0.796** | ### AG-MG | Model | Precision | Recall | F1 | |-------|-----------|--------|------| | UGARIT/grc-alignment (base, t=0.0) | 0.384 | 0.634 | 0.478 | | Dragoman v2 (Homer only, t=0.81) | 0.480 | 0.823 | 0.606 | | **Dragoman v3 + post-processing** (Homer + drama) | **0.680** | **0.948** | **0.792** | v3 evaluated on AG-MG drama gold standard (83 lines across 9 Attic plays, held back from training). Homer gold (Iliad book 2): F1 0.549 (precision 0.387, recall 0.949). ### AER (vs UGARIT gold standard) Raw model output (no post-processing), standard AER (Och & Ney 2003): | | Iliad (148 sent) | Prose (126 sent) | All (274 sent) | |---|---|---|---| | UGARIT base | **19.47%** | 24.71% | 22.59% | | Dragoman | 20.39% | 26.92% | 24.27% | Prose portion (Plato's Crito, Xenophon) held out from training. ### Latest Pipeline Output | Axis | Alignments | |------|-----------| | AG-EN (post-processed) | 118,681 | | AG-MG (post-processed) | 107,223 | ## Training Data | Source | Pairs | Description | |--------|------:|-------------| | EN silver standard | ~15,700 | Passage-level AG-EN pairs from model output + matrix fallback | | MG silver standard | ~15,700 | AG-MG pairs from base model + 7 post-processing heuristics | | Cunliffe lexicon | ~24,100 | AG lemma + EN gloss pairs from [Cunliffe's Homeric dictionary](https://archive.org/details/lexiconofhomeric0000cunl) | | Perseus prose | 20,000 | Sentence-aligned Plato, Xenophon, Herodotus, etc. via [ancient-greek-datasets](https://github.com/kevinkrahn/ancient-greek-datasets) | | NT (OpenGNT) | 26,600 | Clause-aligned Greek NT + English via [OpenGNT](https://github.com/eliranwong/OpenGNT) | | UGARIT gold | 148 | Hand-annotated AG-EN, Iliad portion only (Palladino et al. 2023) | | AG-MG gold | 1,720 | AG-MG pairs from Iliad books 1, 6, 18 (book 2 held for eval) | | Attic drama | 15,800 | Word-aligned AG-MG pairs from Sophocles, Euripides, Aristophanes | Training uses all sources simultaneously with contrastive loss. The model learns cross-lingual representations where aligned words (in any language pair) have similar embeddings at layer 8. ## Post-Processing Pipeline Raw model output is enhanced by a multi-pass post-processing pipeline. These heuristics are applied at inference time, not during training. ### AG-MG (15 passes) 1. **Deduplication** - one AG word per MG word, prefer same-line 2. **Identical/proper noun matching** - surface form or 4+ char prefix match 3. **Lemma matching** - align words sharing any candidate lemma via [Dilemma](https://github.com/ciscoriordan/dilemma) verbose mode 4. **Cross-line lemma fix** - correct misalignments where lemma match exists on a closer line 5. **Reverse cross-line fix** - fix MG words pulled to wrong AG line 6. **Proper noun misalignment fix** - reassign MG proper nouns to correct AG proper nouns 7. **Cross-boundary recovery** - match unaligned words on adjacent lines 8. **Wiktionary AG→MG pull** - AG lemma Greek glosses match unaligned MG words 9. **Wiktionary MG→AG pull** - MG word looked up in Wiktionary, Greek glosses matched against AG (with Katharevousa fallback) 10. **LSJ Greek glosses pull** - AG lemma looked up in lsj.gr Greek-to-Greek definitions 11. **English bridge pull** - AG and MG words sharing an EN gloss via Cunliffe/Wiktionary 12. **Cognate misalignment fix** - correct pairs where a better cognate match exists nearby 13. **Cognate stem matching** - Greek sound-change normalization (β→v, θ→th, αι→ε, etc.) with compound prefix stripping 14. **Epithet equivalences** - Homeric epithet/formula matching (e.g. πόδας ὠκύς → γοργοπόδαρος) 15. **Dependency propagation** - extend alignments along MG dependency arcs ### AG-EN (14 passes + scored matrix) Uses a scored **AlignmentMatrix** where each AG word has ranked EN candidates. Heuristic-added pairs score 3.0, raw model pairs 1.0. Displaced model pairs are suppressed but recover as fallbacks when both sides are unaligned after all passes. Three dictionary sources (Cunliffe, LSJ, Wiktionary) provide gloss-based signal at different score levels, with shared suffix stripping and hyphen-part matching. 0. PhilBerta ensemble - agreement boost (1.5) where Dragoman and [PhilBerta](https://huggingface.co/glaux-nlp/PhilBerta-WordAlignment) agree; PROPN-guarded fallback (0.3) for unaligned AG words 1. Patronymic expansion (Ἀτρεΐδης → "son of Atreus") 2. Compound verb pull (προιάπτω → "sent forth") 3. Case-preposition pull (genitive → "of") 4. Bookending (fill gaps between aligned endpoints) 5. Epithet/vocabulary pull (15 Murray-specific translations) 6. Cunliffe lexicon pull (line-cited definitions) 7. Cunliffe short gloss pull 8. LSJ line-cited pull (8,227 Iliad lines indexed from [lsj9](https://github.com/ciscoriordan/lsj9)) 9. LSJ short gloss pull (117K entries) 10. Wiktionary EN gloss pull (first 5 glosses, skips paradigm noise) 11. Multiword expression pull 12. Same-word propagation 13. Cross-boundary recovery 14. Matrix fallback recovery (suppressed pairs where both sides unaligned) ## Lemmatization Integration Dragoman integrates with [Dilemma](https://github.com/ciscoriordan/dilemma), a Greek lemmatizer covering Modern Greek, Ancient Greek, and Medieval Greek with a 12.5M form lookup table, plus a GPU-optimized POS tagger and dependency parser (96.8% UPOS on AG) available via `dilemma[tagger]`. Dilemma's verbose mode returns multiple candidate lemmas with metadata (language, proper noun status), which the post-processing pipeline uses for cross-matching. The alignment pipeline passes the preceding token as `prev_word` to Dilemma's `lemmatize_verbose()`, enabling article-agreement disambiguation. When a Greek article (e.g. τῆς, τόν, τῶν) precedes an ambiguous form, Dilemma boosts candidates whose gender/number matches the article. This helps distinguish cases like ἔρις (strife) vs Ἔρις (the goddess Strife) based on the syntactic context. Dilemma also supports a `dialect` parameter (e.g. `dialect="ionic"`) for Ionic/Epic forms common in Homer, though the pipeline currently uses the default lookup which already covers most Homeric vocabulary via the AG treebank. The model includes a **lemma disambiguation head** (`lemma_head.pt`) - a small (~200K param) module trained alongside the alignment loss on 52,978 ambiguous targets from the Perseus treebank. It uses Dragoman's contextual embeddings to pick among Dilemma's candidates for ambiguous forms like ἔρις (strife) vs Ἔρις (the goddess). ### Lemmatization Results (Dilemma, standalone) | Benchmark | Tokens | Accuracy | |-----------|-------:|:--------:| | AG Classical (Sextus Empiricus) | 357 | **99.7%** | | Byzantine (DBBE gold standard) | 8,342 | **92.7%** | | Katharevousa (Sathas) | 318 | **95.6%** | | Demotic MG (triantafyllidis convention) | 400 | **96.0%** | | DiGreC treebank (equiv-adjusted) | 118,894 | **93.7%** | ## Usage ### Word alignment with simalign ```python from simalign import SentenceAligner aligner = SentenceAligner( model="open-greek/dragoman", token_type="bpe", matching_methods="a", # argmax device="cuda", layer=8, ) # AG-EN alignment ag = ["μῆνιν", "ἄειδε", "θεὰ", "Πηληϊάδεω", "Ἀχιλῆος"] en = ["Sing", "O", "goddess", "the", "wrath", "of", "Achilles"] result = aligner.get_word_aligns(ag, en) print(result["inter"]) # [(0, 4), (1, 0), (2, 2), (4, 6)] # μῆνιν-wrath, ἄειδε-Sing, θεὰ-goddess, Ἀχιλῆος-Achilles # AG-MG alignment mg = ["Ψάλλε", "θεά", "τον", "τρομερό", "θυμό", "του", "Αχιλλέα"] result = aligner.get_word_aligns(ag, mg) print(result["inter"]) # [(0, 4), (1, 0), (2, 1), (4, 6)] ``` ### With post-processing (full pipeline) ```python # See iliad-align repository for the full pipeline: # https://github.com/ciscoriordan/iliad-align # # python run_alignments.py --axis mg --threshold 0.7 # python align_words.py 1 24 ``` ## Limitations - **Optimized for Homeric and Attic Greek**: trained on Iliad text plus Attic drama (Sophocles, Euripides, Aristophanes). Performance on other AG texts (Plato, Thucydides, NT Greek) is untested. - **Post-processing is Iliad-specific**: patronymic pull, epithet tables, and Cunliffe/LSJ lexicons are Homer-specific heuristics. - **Raw model quality**: without post-processing, alignment F1 is lower (~0.65 for EN, ~0.61 for MG with base model). The post-processing pipeline contributes significantly to the final numbers. - **MG text**: tested only on the Polylas translation (1875, literary Katharevousa/demotic). Modern prose translations may differ. ## Related Work - **UGARIT/grc-alignment** (Palladino et al. 2023): Base model for AG-EN word alignment. Dragoman extends this with AG-MG support, additional training data, and post-processing. - **Dilemma** (Riordan 2026): Greek lemmatizer with 12.5M form lookup + character transformer. Provides lemma matching for alignment post-processing. - **simalign** (Jalili Sabet et al. 2020): Word alignment tool using contextual embeddings. Dragoman uses simalign's argmax algorithm. - **Dilemma tagger** (Riordan 2026, `dilemma[tagger]`): GPU-optimized Greek POS tagger + dependency parser. 96.8% UPOS on AG (Perseus + PROIEL + Gorman), 90.6% on MG. Used for syntactic bonding in redistribution and MG morphological tagging. - **GreTa/PhilTa** (Celano 2025): State-of-the-art AG morphosyntactic parsing. Context-aware lemmatization at 95.6% F1, but AG-only and requires dedicated inference. ## Testing The repo includes a test suite that validates model card metadata, config, tokenizer, weights, and (optionally) end-to-end alignment inference. ```bash # Fast tests only (no model loading, ~3s) python -m pytest tests/ -x -v # All tests including model loading and inference (~10s) python -m pytest tests/ -x -v --slow ``` Requires: `pytest`, `tokenizers`, `torch`, `safetensors`. The slow tests additionally need `transformers` and `simalign`. ## Citation ```bibtex @misc{dragoman2026, title={Dragoman: Diachronic Word...