Ottoman verse no longer needs a scholar's transliteration: full-line scansion accuracy rises from 26.8% to 77.6% on plain modern Turkish spelling

Automatic meter identification and syllable alignment for Ottoman Divan poetry can now start from ordinary modern Turkish spelling, without requiring a scholarly transliteration carrying long-vowel marks and diacritics. The only previously known scanner for Ottoman verse, the deterministic rule-based algorithm of Kurt and Kara, reads syllable weights directly off the letters of a scholarly edition and cannot be used on modern spelling; even on scholarly text, a long-vowel mark the edition omitted is treated as a flaw rather than corrected. On a held-out test split divided by poet, the model scans 77.6% of lines fully correctly, against 26.8% for the rule-based scanner, 55.0% for a neural tagger trained with the same supervision and 63.6% for Claude Opus 5.5 few-shot on the same 500 lines.

The model decides spelling restoration, linking and lengthening choices, and metrical position in one joint search rather than fixing one of them before the rest. Whether a syllable takes a flaw depends on the syllables around it and on the poem's meter, and a single line often fits more than one meter, so syllable-by-syllable decisions cannot settle the reading; systems that fix any one part first perform worse. Replacing the hidden Markov model with a fewest-flaw rule over the same lattice of candidates costs 44.2 points of full-line accuracy; removing the restorer's candidate probabilities costs 49.2 points; removing the two labels for no lengthening and no linking costs 42.4 points. Given the edition's scholarly transcription directly, the model scans 90.6% of lines.

The demonstrated setting is a 52,000-line corpus from thirteen critical editions, split by poet into training, development and test, with evaluation on 500 lines of confirmed scansion annotation; the meter inventory covers fourteen meters, and 124 of the 1,731 test poems fall outside it and are not scored. The tool accepts both plain modern spelling and scholarly transcription with its marks kept, and the code and data pipeline are open-sourced.

The scansion annotation was produced by an ensemble of three large language models under a written guide, with high per-syllable agreement but 66–79% whole-line agreement, and 200 lines were hand-audited; the authors note that a wider test set hand-annotated by multiple experts would help further error analysis. The robustness benchmark is generated from deviation rates measured on web and OCR copies rather than from the real copies themselves, and all models fall to levels unsuitable for real-world use even before the rates are doubled.