data/validation_phonetic_prior/README.md

Phonetic Prior Validation Dataset

Unsegmented inscription fragments for validating the Phonetic Prior algorithm (Luo et al. 2021, "Decipherment of Lost Ancient Scripts as Combinatorial Optimisation").

Designed for use with:

• phonetic-prior-v2/ — improved implementation
• repro_decipher_phonetic_prior/ — faithful reproduction

Purpose

This dataset tests the algorithm's core capabilities:

1. Segmentation: Can it find word boundaries in unsegmented text?
2. Cognate detection: Can it identify cognate pairs from fragmented text?
3. Language discrimination: Does it rank genetically closer languages higher?

Format

Each language directory contains:

{iso}/
  lost.txt              # Unsegmented IPA inscription fragments (one per line)
  known_{other}.txt     # Known IPA vocabulary for each candidate language
  ground_truth_{other}.tsv  # Cognate pairs (lost/known/concept_id)

Plus:

• validation_orderings.yaml — expected closeness orderings
• metadata.json — statistics and provenance

lost.txt

One unsegmented IPA string per line. No word boundaries. Fragments are 1-20 syllables (median ~7), matching Linear A inscription length distribution. ~15% of fragments are cut mid-word to simulate inscription damage.

Example (Oscan):

vibissmintiːsvibissmintiːs
ḍịsfrvernahelvis

known_{lang}.txt

Known-language IPA vocabulary, one word per line. Sampled from data/training/lexicons/{lang}.tsv (up to 2000 items).

ground_truth_{lang}.tsv

Tab-separated: lost, known, concept_id. These are real cognate pairs from cognate_pairs_inherited.tsv.

Languages

ISO	Lost Language	Fragments	Avg Syllables	Family
grc	Ancient Greek	9,217	8.0	IE/Hellenic
lat	Latin	65,017	8.1	IE/Italic
san	Sanskrit	65,614	6.9	IE/Indo-Iranian
ang	Old English	44	5.8	IE/Germanic
osc	Oscan	1,538	5.8	IE/Italic
xum	Umbrian	1,761	6.9	IE/Italic

Validation Orderings

The validation_orderings.yaml file defines expected phylogenetic closeness assertions. Each asserts that when the anchor language is treated as "lost", the closer candidate should score higher than the farther one.

Key test cases:

• Oscan-as-lost: Latin should rank higher than Sanskrit or Greek
• Umbrian-as-lost: Oscan/Latin should rank higher than Sanskrit or Old English
• Latin-as-lost: Oscan should rank higher than Sanskrit

Usage with Phonetic Prior

phonetic-prior-v2

from phonetic_prior_v2.data.adapters import TSVAdapter

# Load as lost text
with open("data/validation_phonetic_prior/osc/lost.txt") as f:
    inscriptions = [line.strip() for line in f]

# Load known vocabulary
with open("data/validation_phonetic_prior/osc/known_lat.txt") as f:
    known_vocab = [line.strip() for line in f]

# Train model on inscriptions + known_vocab
# Score against ground_truth_lat.tsv for P@k metrics

repro_decipher_phonetic_prior

Register as a custom corpus in datasets/registry.py or feed directly to repro/eval/common.py::train_model().

Build

# Step 1: Build segmented inscriptions (if not done)
python scripts/ingest_inscriptions.py
python scripts/build_inscriptions.py

# Step 2: Build unsegmented validation set
python scripts/build_validation_inscriptions.py

Sources

All underlying data is CC BY-SA 4.0 compatible. See data/inscriptions/README.md for full source listing.