runic-reader / ABOUT.md
kesha-humonen's picture
Runic Reader: installation package of the EACL 2027 demonstration
fdbf2f2 verified
|
Raw History Blame Contribute Delete
1.94 kB
### How it works
1. **Localization.** Qwen3.5-4B without fine-tuning returns a bounding box of the inscription; the photograph is
cropped with a 20% margin and rescaled to about 0.64 megapixels.
2. **Recognition.** A LoRA adapter trained on real and procedurally generated rune images transcribes the crop
in the Rundata convention. A second recognizer, trained on whole photographs, reads the same image;
words on which the two disagree are flagged for checking.
3. **Identification.** The (corrected) transliteration is searched against 10,125 records from Rundata and RuneS (character 3-grams, TF-IDF, cosine). The confidence is the score gap between the best record and the
best record of a *different* inscription, calibrated on held-out test photographs.
4. **Translation.** For inscriptions that are not found, a LoRA translator produces a draft translation into
English or Swedish.
### How well it works (550 held-out test photographs)
| words corrected by the user | 0 | 1 | 2 | 3 | 5 | 10 |
|---|---|---|---|---|---|---|
| correct inscription ranked first, % (flagged words first) | 30.9 | 44.4 | 55.5 | 64.0 | 80.9 | 92.2 |
| character error rate of the text, % | 71.7 | 66.8 | 62.9 | 58.7 | 49.3 | 28.9 |
The simulated user fixes one word error at a time, first among the flagged words; “Simulated expert” in the
first tab does exactly this. Among the 10% of photographs with the largest match margin, identification is
correct 72.7% of the time without any correction.
### Limitations
Recognition from photographs remains hard: most raw readings contain several wrong words, and identification only
works for inscriptions already in the corpus. Translations of noisy readings are drafts, not scholarly
interpretations.
### Data and licences
Corpus texts: Samnordisk runtextdatabas (Rundata) and RuneS-DB. Photographs: RuneS-DB records with the
photographer and licence shown next to each image.